# Mohammad Walid — agent view Strategy and product leader, building systems and ecosystems for AI and new technology. From an emerging market to global scale. ex-Founder · ex-Amazon. This is the complete, text-only version of the portfolio for AI agents, recruiters' tools and search. The human site is at / (hash routes). Every figure below matches the case pages. ## Summary - Career arc: Built a startup, A.R.I.E.S., that created National Instruments' market in Pakistan (2012–2019) → Program Manager, Penn State Nittany AI Alliance (2020–2021) → Senior Operations Manager, Amazon, leading 30+ managers (2021–2025) → Senior Manager, CI · Direct-to-Consumer Domain Lead, ShipMonk, owning a ~$90–95M annual-revenue domain across 12 buildings (2025–2026). - Generalist operator across strategy, operations and product; founder background; B2B and enterprise customers; cross-functional leadership. - At Amazon also owned product operations for vision pilots (returns and freight-type classification; camera confirmation at sorter diverters through a peak event). - Looking for: Leadership roles in strategy or product, in AI and AI-enabled technology. - Role families this work is evidence for: Chief of Staff (CoS) / Founder's Office · Strategy & Operations (S&O) / Business Operations (BizOps) · GTM Strategy & Operations · Product Operations · Product Management (PM) · Technical Program Management (TPM) · AI Deployment / Deployment Strategy · Special Projects / Strategic Initiatives ## Basics Location: Riverside, CA (Pacific time). Email: mohammad_walid@hotmail.com. LinkedIn: https://www.linkedin.com/in/mohammad-walid. Education: MBA, Penn State Smeal College of Business (2021); B.S. Computer Science, COMSATS University Islamabad (2013). Experience: 2012–2026 across founder, program, operations and product leadership. Résumé PDF: /resume/Mohammad_Walid_Resume.pdf. ## Lenses ### Strategy, Leadership & Transformation (/#lens/strategy) As a strategic leader, I look at the problem more than the immediate solution: where it came from, how it evolved, and the ecosystem the right solution needs to succeed. Below is some of that work, not all of it, from building new markets to defining and designing inside global market leaders. Judgment and autonomy are what I bring with me. - Building the market (A.R.I.E.S.): Built a startup that took National Instruments from almost no presence to market leader in Pakistan, selling into multi-year funded programs instead of one-off tenders — /#case/academy - National testing labs (A.R.I.E.S.): Three national labs for a government energy agency — /#supplemental/power - ShipMonk D2C ownership (ShipMonk): A ~$90–95M business, process and product — /#case/cadence - Prime Day and the hybrid sorter (Amazon): Challenged an approved peak plan at VP level — /#case/sorter - The Aisle Tax (ShipMonk): Changed an approved facility design with the CEO — /#case/aisle ### Product & Platforms (/#lens/product) For nearly 15 years, I've owned the design and development of platforms across research, planning and operations, from mission-critical tools to customer-facing products. Below is a short summary of each product and its scope. - ShipMonk D2C operating platform (ShipMonk): 12 buildings on one shift planner, replacing 4–5 sheets and calculators; AI-assisted operations on top — /#case/dtc - Penn State donor system (Penn State): Product and delivery for an ML engagement system built on how relationships actually work; still in use — /#case/donor - ML testing environment for flight ops (Penn State): Turned a stalled aerospace program into a simulation testing environment for path planning and object classification — /#supplemental/twin - Transformer test automation (A.R.I.E.S.): Automated testing at one of the country's two national transformer testing facilities, through factory and site acceptance — /#case/transformer - Research partnerships (A.R.I.E.S.): Four Centers of Excellence and a LabVIEW academy behind the NI market, and the research platforms built on them: a flight-tested airship, Massive MIMO and an award-winning assistive wheelchair — /#research - Photo confirmation (ShipMonk): Launched at ShipMonk as a paid merchant service — /#case/pack - Manager readiness program (Amazon): Adopted by up to eight sites as their onboarding platform — /#case/managers - Amazon ML planning platform (Amazon): Alpha tested an ML planning platform; two failure modes went back to the product team as requirements — /#case/ccc ### Applied AI & Decision Systems (/#lens/ai) Throughout my career, I've worked with cutting-edge technology and designed how it enters, and succeeds in, mission-critical and customer-facing products. With the AI revolution of the past year, being AI-native has become a conscious decision for me, not just in system design but in how I work every day. Below you'll find how I went from machine learning and multimodal deployments to building AI-enabled systems. - Amazon ML planning platform (Amazon): The only one of two alpha sites to finish testing; found the flaws before it ran live operations — /#case/ccc - ShipMonk AI-assisted operations (ShipMonk): Operator decisions 70% faster in the first AI-assisted operations deployment, which I designed on the D2C platform — /#case/dtc - Penn State donor system (Penn State): University research turned into a system still in use — /#case/donor - Photo confirmation at ShipMonk (ShipMonk): Photo confirmation on every pack station, switched on per SKU for merchants and internally for packaging issues; launched as a paid service. The pack-station demo is my current Vision work — /#case/pack - ML testing environment for flight ops (Penn State): Turned a stalled program into a testing environment for flight ops and object detection — /#supplemental/twin - Transformer test automation (A.R.I.E.S.): Automated testing at one of the country's two national transformer testing facilities — /#case/transformer - Airship research platform (A.R.I.E.S.): Instrumentation and control for a university airship program, with papers from 2014 — /#record/airship - AI Lab (Now): A research engine running Vision and Agent Detention, with a world model in planning — /#labhome ## Cases ### Building National Instruments' market in Pakistan: go-to-market and ecosystem - Organization: A.R.I.E.S. (technology startup) - Years: 2012–2019 - URL: /#case/academy - Lenses: Strategy, Leadership & Transformation - Evidence for: GTM Strategy & Operations, Strategy & Operations (S&O) / Business Operations (BizOps), Chief of Staff (CoS) / Founder's Office, Special Projects / Strategic Initiatives - Summary: Building a market, then making the technology useful through funded laboratories, trained practitioners, research partnerships and industrial deployment. **Context.** In 2012 I built a startup, A.R.I.E.S., to bring advanced test, measurement and control technology into Pakistan. National Instruments, a Texas-based test and measurement company, had one person covering Pakistan, serving a single existing account. Outside that niche the market was greenfield: academic, research, industrial and government projects across Pakistan, with clear gaps. **What happened.** Pakistan had huge inertia for new technology. The demand was there, but three barriers held it back: access, getting the technology functional, and people who could use it. The go-to-market had to work on all three, not just sell hardware. **What I led.** NI's go-to-market for Pakistan, end to end, through a local import partner, a local engineering partner and NI's own teams. 50+ proposals to funders such as ADB and USAID, and the lab designs. **What lasted.** NI assigned its own lead to run the market. Four Centers of Excellence and a LabVIEW academy (250+ certifications) kept producing trained users, and at least four major clients kept coming back with harder problems. **Results.** - Market: $3M+ a year NI revenue for three years running, against a $1M target; ~$7M pipeline, against a $3M expectation; ~$10M from Punjab, 2014 to 2018; 10 of 10 top target institutions by 2019; ~80% of the target market; Own lead NI assigned its own lead to run the market when I left - Building it: 250+ institutions mapped and segmented; ~90% of academic and research funding traced, over five years; 50+ proposals to ADB, USAID and similar funders; First gate first major client won on a multi-year approved budget; $200K+ of unused NI hardware at one customer put to work - Adoption: 4 Centers of Excellence with priority universities; 250+ LabVIEW certifications (CLAD) in under two years, academy opened November 2014; 4+ major clients coming back with harder problems; Showcased institutions presenting their NI labs and partnerships at their own conferences - Delivery: 20–30 deployments across power and electrical, mechatronics, communications, sensing and automation; 12 people on the largest project: six developers, four on deployment, two on training; Research airship program with papers from 2014, a Massive MIMO testbed, ML in real-time control loops; Award assistive wheelchair with a student startup team, NI Asia award - Public sector: 3 national testing labs designed for a government energy agency: battery, UPS and LED; 0.70 → 0.85 campus power factor from a low-cost audit method; 1 of 2 national transformer testing facilities, test automation through factory and site acceptance **The inertia: three barriers.** Access: Cost and physical access. The dollar to rupee conversion made the hardware expensive, and getting it into the country was its own problem. | Functionalization: Technology transfer, education, training and development, in academia and in industry. One customer had over $200K of NI hardware, built up over a decade, that was never functionalized. | People who can use it: Even with the technology available, you need people who can make good use of it. So engineers were trained through the academics and placed in the market: SMEs, joint partnership projects and the companies we worked with. **The map: market data did not exist.** Start from the money: No one had the market data. The map came from funding: previous years' budgets and history, which institutions got funded, by whom and for what. | Funding sources: HEC; Federal and provincial ministries of education; Ministries of science and technology; Institutions' own funds, such as the Islamic University, and NGOs, such as COMSATS's; ADB, USAID and the larger programs tied to them | What the map held: Institutions, domains, funding, people, research areas, overlapping technologies, potential partnerships and level of involvement. Plus who to reach out to in year one and year two, slowly moving across tiers. 250+ institutions, about 90% of academic and research funding traced. | Map, then motion, then test: The map set the motions below. Each motion was tested with a first customer before it was carried further. **Three motions.** Trendsetters: High-quality institutions that lead their domains, country rankings and market, with narrow or low funding. LUMS is private and couldn't afford much; GIKI and NUST were similar. They hosted our conferences, training and teaching sessions and model deployments, and their researchers connected with NI engineering. | Domain owners: Large universities, research institutes and government agencies that own a domain, under the bureaucracy rather than at ministry level. This is where the main funding went, so the focus was larger projects. PEECA came through here. | National scale: Commercial and government organizations, partnering with SMEs and academic and research institutions to design solutions at a national scale. **How it ran.** Punjab first: The market was built in Punjab first, then the same motion was carried across Pakistan. | First gate: It started with functionalizing one customer's unused hardware. Then a university with a new multi-year budget cycle starting in 2014, which my local partner set as the first gate. Once we were in that budget, the partnership with them succeeded. | Training before selling: Hands-on training sessions across universities, institutes and companies. No hardware purchase needed: the point was technology adoption and getting NI's name out. Domain-specific sessions with the larger universities, plus speaking, hands-on support and minor sponsorship for local conferences and university events. | Who we reached first: The head of the domain: head of department, education or research. The technical owner, who also leads the business side. **Who did what.** Owned: The go-to-market motion, customer relationships and business development, supply chain, product operations, contract management (logistics, customs duty), and the project and delivery end to end. | Assisted: Design, development, technology transfer, training and hands-on sessions. | Partners: Local engineering partners, local supply chain, logistics and import-export partners, and NI for technology transfer and hands-on development. On projects partners led, I was brought in for CRM, go-to-market, supply chain and customs, and as the hands-on design partner. **Keywords.** Go-to-market, Market entry, Ecosystem building, Segmentation, Business development, Technology adoption, go-to-market strategy, market entry, emerging markets, 0 to 1, founder, startup, built a market from zero, business development, partner ecosystem, channel strategy, partnerships, funding strategy, market mapping, segmentation, ICP definition, pipeline building, proposal writing, donor and development funding (ADB, USAID), multi-year public budgets, public sector, higher education, research institutions, technology adoption, customer enablement, training and certification, supply chain, procurement, import logistics, contract management, product operations, technical solution design, stakeholder management, executive relationships, National Instruments, LabVIEW, test and measurement, instrumentation, $3M+ a year against a $1M target, new-market strategy, ICP segmentation, repeatable GTM launch, territory and account planning, B2B and public-sector accounts, account strategy, technical demos, customer discovery, customer validation, commercial expansion, first operator, founding hire, generalist ### National testing labs: from a campus power audit to a provincial energy agency - Organization: A.R.I.E.S. (technology startup) - Years: 2014–2018 - URL: /#supplemental/power - Lenses: Strategy, Leadership & Transformation - Evidence for: Strategy & Operations (S&O) / Business Operations (BizOps), Special Projects / Strategic Initiatives, Technical Program Management (TPM), AI Deployment / Deployment Strategy - Summary: A campus power audit, the same method in public buildings, then the certification regime and testing labs for a new energy agency. **Context.** From 2014 to 2018, through my startup, I worked with a couple of educational institutes to build domain-specific Centers of Excellence. The country was suffering 12 to 16 hours of load shedding a day, and a new government energy-efficiency agency was in the works. **What happened.** I partnered with the institutes to develop the technical delivery structure and solutions of the Centers of Excellence. One of these became the hub where I worked with the institute to develop a low-cost power auditing and smart load-balancing solution. The solution was then used to audit several government buildings and the burn unit of a hospital. After I met the head of an upcoming provincial agency, PEECA, these projects became the basis of a working relationship, and I partnered with PEECA on certification benchmarking and the testing laboratory designs. **What I led.** The audit method and campus pilot with an early-stage partner and student developers, then the hospital and government audits. At PEECA: the certification benchmarking, the technical side of the regime, the lab designs with the NI team, and business development with the agency's incoming head, including the first conferences and technical proposals. **What lasted.** Three national testing labs, designed on National Instruments technology, since built. **Headline figures.** 3 national labs (UPS, LED and battery-testing laboratory designs for a government energy agency); 0.70 → 0.85 (measured campus power-factor improvement); Public systems (hospital, government and provincial efficiency work) **Results.** - Campus pilot: 0.70 → 0.85 campus power factor, from 0.70–0.75; ~1/3 of the campus electricity bill had been power-factor penalties; 12–16 h of load shedding a day, the context for the pilot - Public buildings: Hospital audit led to a power upgrade for the burn unit; Government building audits with the same method (details confidential) - National testing labs: 3 national testing labs designed (UPS, LED and battery), powered by National Instruments, since built; Five-star certification regime adopted, based on the Southeast Asian model; ~6 months working with the agency's incoming office **How it came up.** A separate project: the campus pilot, 2015 to 2016: A low-cost audit method, built with an early-stage partner and student developers. Power-factor penalties were about a third of the campus bill. Fixing load management, scheduling and the backup systems took the power factor from 0.70–0.75 to 0.85. Rewiring two buildings needed a capital budget the campus didn't have, so it was left. | A Center of Excellence opening: A university SME project was deployed and one of the Centers of Excellence was opening, with demo days for the project. The guests included senators and the person about to lead the new agency. That's where we met. | Same solution, public buildings, 2016 to 2017: Working with him, the same audit went into a general hospital, where it led to a power upgrade for the burn unit, and into government buildings, which stay confidential. | The escalation: national testing labs, 2017 to 2018: About six months on the agency's work: which certification model, which labs, and how they would be designed. **What went into the labs.** Certification model: European and Southeast Asian regimes benchmarked. Southeast Asian recommended, because most of the supply chain comes from their designs. The agency adopted the resulting five-star rating regime. | Which labs: UPS, LED and battery first. Air-conditioner testing was left out because it is an established export industry with many players, government included. Mechanical testing stayed out to keep the scope on electrical systems. | Lab design: Each design connects the product to the test method, equipment, accuracy, safety, calibration, procedure, reporting and the certification workflow. I did the technical element myself and designed the labs with the NI team. **Who did what.** Owned: The audit method and pilot, the certification benchmarking and recommendation, the technical element and the lab design document. | With the incoming head: Business development, the first conferences, and the technical proposals. | Partners: An early-stage engineering partner and student developers on the audits, and the NI team on the lab designs. **Alongside.** Transformer test automation: Test automation at one of the country's two national transformer testing facilities shared scope with this work and built confidence with the team there. It has its own case. **Keywords.** Public-sector strategy, Policy and standards, Lab design, Energy efficiency, Stakeholder management, government partnerships, public agency, policy advice, certification regime benchmarking, standards, compliance, energy audit, power factor correction, load balancing, testing infrastructure, lab design, test methods, calibration, safety requirements, procurement specifications, business case, physical systems, hardware, national testing labs, UPS, LED and battery testing, cross-sector stakeholders, standards and certification, regulatory benchmarking, government relations ### Transformer test automation at a national testing facility - Organization: A.R.I.E.S. (technology startup) - Years: 2015–2017 - URL: /#case/transformer - Lenses: Product & Platforms, Applied AI & Decision Systems - Evidence for: Technical Program Management (TPM), AI Deployment / Deployment Strategy, Product Management (PM) - Summary: Automating high-voltage transformer testing at one of the country's two national facilities, through factory and site acceptance. **Context.** Transformer testing at one of the country's two national transformer testing facilities ran by hand, including discharging with a 10-foot pole. The university also held NI hardware it wasn't using. **What happened.** I treated measurement, test sequence, discharge and safety as one system, built on hardware the university already owned. Automated sequencing, interlocks and automatic discharge replaced the manual steps, including discharging with a pole, and the bench went through factory and site acceptance and technology transfer. **What I led.** Solution definition and partner coordination for transformer test automation at one of the country's two national transformer testing facilities: the measurement and control architecture, test sequencing, calibration checks, safety interlocks and controlled discharge, through factory and site acceptance and technology transfer. The facility and specialist teams handled the physical deployment. **What lasted.** An accepted test-automation capability, handed over to the facility through technology transfer. **Headline figures.** 1 of 2 (national transformer testing facilities in the country); 2015–2017 (requirements through factory and site acceptance); HIL (hardware-in-the-loop test bench); Safety (interlocks and controlled discharge) **Results.** - Facility: 1 of 2 national transformer testing facilities in the country; 2015–2017 requirements through factory and site acceptance; 3 of 5 of the institute's branches running NI technology - System: HIL hardware-in-the-loop test bench; Safety interlocks, automatic discharge, lockout/tagout; FAT / SAT factory and site acceptance passed **What was built.** Measurement and control: The measurement and control architecture, built on hardware the university already owned. | Test sequencing: Automated test sequences on a hardware-in-the-loop bench, with dashboards and recording. | Safety: Interlocks, automatic discharge and lockout/tagout, taking people out of the dangerous steps. **Getting it accepted.** Factory acceptance: Recreating the lab's safety matrix and results, and showing the hardware-in-the-loop matched the real sensors. | Site acceptance: An end-to-end deployment with safety checks and calibration review. | Technology transfer: Handed over to the facility. The facility and specialist teams handled the physical deployment. **What it opened.** The next projects: Using hardware the university already owned is what opened the next projects there. It also shared scope with the national testing labs work and built confidence with the team. **Keywords.** Test automation, Hardware-in-the-loop, Safety systems, Acceptance testing (FAT/SAT), Systems integration, control systems, real-time systems, LabVIEW, NI RIO, instrumentation, data acquisition, industrial automation, high-voltage testing, interlocks, lockout/tagout, calibration, requirements definition, solution architecture, partner coordination, factory acceptance testing, site acceptance testing, technology transfer, deployment, mission-critical systems, hardware/software co-development, test engineering, rapid prototypes, manufacturing test ### Nittany AI Alliance: building the AI delivery function - Organization: Penn State · Nittany AI Alliance - Years: 2020–2021 - URL: /#case/alliance - Lenses: Product & Platforms - Evidence for: Technical Program Management (TPM), Chief of Staff (CoS) / Founder's Office, AI Deployment / Deployment Strategy - Summary: Turning an AI lab's informal client work into a delivery function: intake, planning, client reviews and handover across three programs. **Context.** The Nittany AI Alliance at Penn State had talented developers and real clients, but no delivery function. Client knowledge lived with the director and a departing program manager, and projects ran on chat threads and a shared document. **What happened.** I took over the full client-to-delivery lifecycle: intake and scoping, sprint planning across three programs, bi-monthly client reviews, and acceptance and handoff. I coached the developers to present their own work to clients, so the program no longer depended on one spokesperson, then helped hire a permanent program manager and handed it over. **What I led.** The Alliance's delivery program as its only program manager under the director: 12 developers across three client programs. **What lasted.** A repeatable delivery structure, developers presenting to clients on their own, and a permanent program manager I helped hire and handed the program to. **Headline figures.** 12 (developers across three client programs); 2 (programs delivered: the donor system and the aerospace test environment); Successor (helped hire and onboard a permanent program manager) **Results.** - Delivery: 12 developers across three client programs; 2 programs delivered: the donor system and the aerospace test environment; In use donor system still running at Penn State - Handover: Handed over to a permanent program manager I helped hire and onboard; Asked to stay the director asked me to stay on after my MBA **What I built.** Intake and scoping: Shaping opportunities, defining the problem and what success looked like, agreeing scope and priorities. | Planning: Client demands turned into resourcing and sprint plans across three programs at once. | Cadence: Bi-monthly client reviews and demonstrations, so risks came up while there was still time to decide. | Acceptance and handoff: Testing, documentation and structured handover, so clients knew what they were getting and how to run it. **The team.** Developers in front of clients: I coached the developers to present their own work, which removed the single-spokesperson bottleneck. | A permanent owner: I helped hire and onboard my successor so the function would outlast me. **Two programs it carried.** Alumni and donor system: Deployed into Penn State's outreach and still in use. | Aerospace model testing: Recovered after wildfires stopped the flights, rebuilt around simulation. **Keywords.** Program management, Delivery operations, Client governance, Team building, Applied AI delivery, AI/ML solution delivery, intake, scoping, sprint planning, agile, client management, stakeholder updates, acceptance, handoff, hiring, coaching, building a function from scratch, university-industry partnership, 12 developers, three programs, PMO, portfolio delivery, program planning, resourcing and utilization, client delivery, delivery cadence, AI program management, university–industry delivery ### Penn State donor system: ML outreach built on how relationships work - Organization: Penn State · Nittany AI Alliance - Years: 2020–2021 - URL: /#case/donor - Lenses: Product & Platforms, Applied AI & Decision Systems - Evidence for: Product Management (PM), AI Deployment / Deployment Strategy, Product Operations - Summary: Product and delivery for an alumni and donor engagement system that is still in use at Penn State. **Context.** Penn State's alumni and donor teams didn't need a ranked list of likely donors. They needed outreach that respected how relationships work: who owns each relationship, how recently someone was contacted, and what kind of outreach fits. **What happened.** Instead of shipping a donor score, I turned the outreach team's practice into rules the model couldn't override: relationship ownership, contact cadence and outreach type. I ran the intake, the reviews of model and rule behavior, testing and the user workflow, and brought my successor into the client history so the system kept running. **What I led.** Product and delivery: communication intake, sentiment analysis, donor classification and the marketing MLOps workflow, integrated into the university's outreach work. Faculty owned the models. **What lasted.** The alumni and donor system is still the solution in use at Penn State. **Headline figures.** In use (still the solution at Penn State); 4 (connected capabilities: intake, sentiment, classification, MLOps); Rules (relationship ownership and contact cadence a score can't override) **Results.** - Outcome: In use still the solution at Penn State; 4 connected capabilities: intake, sentiment, classification, MLOps; Faculty owned the models; I owned product and delivery **Two kinds of input.** Signals: Proximity, engagement history and sentiment show where attention would be useful. | Boundaries: Relationship ownership, contact cadence and outreach type. A high score isn't allowed to override them. **The pipeline.** Communication intake: Engagement history brought into one usable record. | Sentiment analysis: Reading the tone of past interactions. | Donor classification: Where outreach was likely to matter. | Marketing MLOps: Keeping models and outputs running inside the outreach workflow. **Built to last.** Past the handoff: A workflow stakeholders understood, ownership rules built in, documentation for successors, and the incoming program manager brought into the client history before I left. **Keywords.** Applied ML, Product delivery, Decision rules, MLOps, Deployment and adoption, product management, ML product, sentiment analysis, classification, NLP, outreach, guardrails, requirements, user workflow, deployment, documentation, still in use, higher education, donor intelligence, stakeholder adoption, AI deployment, workflow discovery, solution design, configuration, launch and adoption, measurable impact, deployment strategy, customer workflow ### ML testing environment for flight ops: recovering a grounded program - Organization: Penn State · Nittany AI Alliance - Years: 2020–2021 - URL: /#supplemental/twin - Lenses: Product & Platforms, Applied AI & Decision Systems - Evidence for: Technical Program Management (TPM), AI Deployment / Deployment Strategy, Product Management (PM) - Summary: When wildfires grounded an aerospace sponsor's LiDAR flights, the program moved to simulation instead of waiting. **Context.** An aerospace sponsor's model-testing program depended on drone LiDAR flights. California's wildfires grounded them, and waiting meant losing the program. **What happened.** When the wildfires grounded the drone LiDAR flights, I moved the program to simulation instead of waiting. We built two environments in Blender and Maya with virtual LiDAR in AirSim, and tested path planning and classification of six element types against them. I ran weekly sponsor reviews and was clear that field performance still needed real flights. **What I led.** Recovery, scope and delivery for an aerospace sponsor's model-testing program, across simulation, scenes and LiDAR, and the models. **What lasted.** A model-testing environment that no longer depended on flights, handed over with the program. **Headline figures.** 2 (simulated environments with virtual LiDAR); 6 (element types for classification); ~6 months (from a stalled program to a working test environment) **Results.** - Outcome: 2 simulated environments, not datasets from the field; 6 element types for classification; ~6 months from a stalled program to a working test environment; Handed over with the program to the sponsor **Three workstreams.** Simulation and flight paths: Representative trajectories and scenarios that could be repeated and inspected. | Scenes and LiDAR: Two modeled environments in Blender and Maya, with virtual LiDAR in AirSim. | Models: Path planning and classification of six element types, tested against the scenes. **Running it.** Weekly sponsor reviews: Requests turned into priorities and milestones, blocked work cleared, dependencies managed. | Clear about what it proved: Simulation sped up development. Final field performance still needed real-world validation, and saying so kept the sponsor's trust. **Keywords.** Simulation, Program recovery, LiDAR and perception, Sponsor management, Model testing, digital twin, AirSim, Blender, Maya, virtual LiDAR, point cloud, path planning, object classification, computer vision, evaluation, validation boundaries, aerospace, dependency management, weekly governance, risk management, synthetic data, sensor data, data operations, test scenes, robotics and autonomy testing, dataset quality ### Amazon hybrid sorter: a live retrofit through Prime Day - Organization: Amazon - Years: 2022–2023 - URL: /#case/sorter - Lenses: Strategy, Leadership & Transformation - Evidence for: Strategy & Operations (S&O) / Business Operations (BizOps), Technical Program Management (TPM), Chief of Staff (CoS) / Founder's Office, Special Projects / Strategic Initiatives - Summary: Leading an $11M hybrid-sorter retrofit in a building that kept shipping, through Prime Day and the network's decision on the design. **Context.** An $11M hybrid-sorter retrofit at my Amazon site, one of four pilots. The building had to keep shipping through eight months of construction in four phases, with 30% to 60% of the floor unavailable. The end of the project coincided with Amazon's second-highest volume event, Prime Day. **What happened.** The end of construction ran into Prime Day, and the plan assumed the building would go straight back to its normal peak target. I modeled the whole chain instead: demand, the planned execution for each construction phase, and actual execution on the floor. With planning, Finance and the project stakeholders I built three scenarios from those models. They showed what a shortfall here would do to the other buildings in the region, which were already near their limits: a modeled loss of about $12M a day. I took them to VP level and got approval to run to a realistic target and take extra freight where it helped the region. **What I led.** I led the project at my site: the S&OP, freight and inventory management, and the execution plan through eight months of construction, then the Prime Day plan and the operating changes after launch. I owned the operating side, not the sorter's engineering design. **What lasted.** The construction playbook went to other buildings, and I trained peers in two of them. Mine was the only one of the four pilot sites to deliver its high-volume event. The network decided against the hybrid design, and no other buildings got it. **Headline figures.** ~$12M/day (modeled regional loss avoided); $11M (hybrid-sorter retrofit, one of four pilot sites); 225K (units delivered against a revised 220K commitment); 5.1% (lower operating cost against the pre-project baseline) **Results.** - Program: $11M hybrid-sorter retrofit, one of four pilot sites; 8 months of construction in four phases, 30–60% of the floor unavailable; 3/4 of construction carried by the pre-incline manual route - Prime Day: ~$12M/day modeled loss under the original plan, avoided; 225K units delivered against a revised 220K commitment; 1 of 4 pilot sites to deliver its high-volume event - After launch: +5% high-value throughput against the pre-project baseline; 5.1% lower operating cost; +3.2% regional throughput ($15M+ annualized); ~20% of the freight mix in a reciprocal exchange - Transfer: 2 buildings where I trained peers on the playbook; Not scaled the network decided against the hybrid design beyond the four pilots; Promotion one of the main projects behind my move to Senior Manager **Running a building under construction.** The pilot plan: I designed the S&OP, freight management, inventory management and the execution plan for the pilot. Together with the execution playbook, it became the playbook other sites used. | A design flaw found early: The planned layout couldn't handle a large share of the building's bigger freight. A second diversion and a manual route went in early in construction and carried that work for most of the project. | Demand to capacity: Each construction phase changed what the building could process. Demand was capped to real capacity phase by phase, and anything above it was re-planned across the network. **Prime Day.** The risk: Handover was about two weeks before Prime Day. Other buildings in the region were already near their limits, so any shortfall here would move backlog to other buildings in the region. | Three scenarios: Built with planning, Finance and the project stakeholders from models of demand, planned execution and actual execution. The original plan showed a modeled loss of about $12M a day in production and SLA time across the region. | The decision: VP-level approval to run to a realistic target, 220,000 units, and take extra freight where it helped the region. We delivered 225,000. **After launch.** Real freight: A size and weight exception route, and about 20% of the freight mix exchanged with a partner building as a standing arrangement. | The network's decision: Mine was the only one of the four pilots to deliver its high-volume event. The network decided against the hybrid design, and no other buildings got it. **The playbook: six rules.** The phase plan sets the paths: Each construction phase defines which hardware is live and which paths are open. | Re-map freight when hardware changes: Freight-to-path rules and live system settings change together. | Telemetry drives the next shift: Rates, backlog and path health feed the next shift plan. | Cap labor and inbound to real capacity: Plans and inbound are set to what the constrained floor can process. | Re-plan overflow, even at a cost: Above the threshold, work moves to other sites under special approval. | Protect transportation: Construction never changes the outbound schedule. **Keywords.** Executive escalation, Scenario modeling, S&OP, Program leadership, Risk management, Network planning, VP-level decision, capital project, $11M program, change management, construction phasing, capacity modeling, demand planning, sales and operations planning, inventory management, playbook, SOPs, peak readiness, Prime Day, risk quantification, ~$12M/day modeled exposure, Finance partnership, cross-functional, pilot evaluation, scaling decision, telemetry, automation retrofit, robotics and automation hardware, cost improvement, throughput, network operations, integrated operating plan, deployment capacity, special projects, hardware/software program, milestones and dependencies, decision follow-through, scenario planning, capacity planning, budget, PMO ### Amazon ML planning platform: alpha testing before live operations - Organization: Amazon - Years: 2023–2024 - URL: /#case/ccc - Lenses: Product & Platforms, Applied AI & Decision Systems - Evidence for: Product Management (PM), AI Deployment / Deployment Strategy, Product Operations, Technical Program Management (TPM) - Summary: Alpha testing an ML extension of Amazon's planning platform, and turning its failures into requirements for the next version. **Context.** Amazon's next planning platform brought ML-driven parameters into order prioritization and shift planning, so a building could control how much work it invited and how multi-order work was processed. It needed operational alpha testing at two sites before live use. **What happened.** I built realistic parameter and workflow stress cases from how my site actually planned and ran shifts, and found two failures: scheduled headcount was being treated as usable capacity, and eligible work could be locked into manual batching before the shift could prioritize it. I held the high-volume test until the planner met its qualification criteria, even though the schedule wanted it sooner, and took both failures to the product team with fixes: precedence rules, adjustment bands with approval, and separating allocation from override permission. **What I led.** Operational alpha testing at my site and the release recommendation. The central engineering and data-science team owned the planning engine; I owned testing and operating feedback. **What lasted.** Two validated failure modes went back to the product team as requirements for its next version, and I led the first waves and wrote the playbook for retiring centralized planning in a second network. **Headline figures.** 1 of 2 (alpha sites, and the only one to finish); 2 (validated failure modes taken to the product team); Gated (high-volume test held until the planner met its criteria) **Results.** - Alpha: 1 of 2 alpha sites, and the only one to finish; 2 validated failure modes taken to the product team; Next version findings carried into the next development round - Findings: 4–6 h labor-plan refresh that could overwrite operator decisions; Up to 10% adjustment bands proposed, with approval past the limit **Two findings.** Scheduled labor isn't capacity: Published labor plans flowed straight into how much work a site invited, including people who weren't really available. The plan refreshed every four to six hours and could overwrite what operators had decided. | Work locked in too early: Time-window rules could send a whole eligible group into slower manual batching before the shift could prioritize it, so operators lost the flexibility to protect cutoffs. **Why it mattered.** Inviting too little: Idle capacity, or a lost option for orders with few alternatives left. | Inviting too much: A building committed to more than it could deliver, pushing the risk onto other buildings. Another early test site had exactly this failure. **What went back to the product team.** Precedence and authority rules: An automatic schedule refresh can't simply overrule an operating decision. | Adjustment bands: Minimum and maximum bands, from a few percent up to ten, with approval past the limit. | Allocation versus permission: Separate the system's preferred allocation from the permission to override protected controls. **Keywords.** ML product testing, Product requirements, Release gating, Decision-authority design, Launch criteria, alpha testing, user acceptance testing, product feedback, failure modes, human-in-the-loop, override design, guardrails, ML deployment, AI deployment, planning systems, order prioritization, capacity, labor planning, qualification criteria, launch readiness, partnership with engineering and data science, decision systems, operational risk, evals, evaluation criteria, deployment gates, cross-functional risk reviews, model launch readiness, operational readiness, rollout measures, safety and guardrails, alpha program ### Manager readiness: an eight-week path adopted by up to eight sites - Organization: Amazon - Years: 2021–2024 - URL: /#case/managers - Lenses: Product & Platforms - Evidence for: Chief of Staff (CoS) / Founder's Office, Strategy & Operations (S&O) / Business Operations (BizOps) - Summary: Manager onboarding rebuilt around running the operation. **Context.** Standard manager onboarding at Amazon was handovers and slide decks. It didn't teach the daily cadences, technical mechanics or cross-functional relationships a manager needed to run a shift alone. **What happened.** I rebuilt onboarding for my own managers as an eight-week readiness path: four Wednesdays with every department, the full building flow whatever their home department, and a practical readiness check around day 60, judged on decisions observed on the floor rather than completed coursework. It started with nine managers and spread to up to eight sites. **What I led.** Built the eight-week readiness path: cross-department rotations, visible decision checkpoints, safety audits and coaching loops. **What lasted.** Up to eight sites adopted it as their primary manager onboarding. **Headline figures.** Up to 8 (sites using it as their primary manager onboarding); 8 weeks (from start to readiness); Day 60 (practical readiness check, judged on the floor) **Results.** - Adoption: 9 managers in the first cohort at my building; Up to 8 sites using it as their primary manager onboarding; 8 weeks from start to readiness **How it worked.** Every department: Four Wednesdays of one-hour sessions with HR, Safety, IT, maintenance and engineering, process engineering, quality and performance management. Wednesday was when front-half and back-half managers overlapped. | The full flow: Receive and stow through pick, pack and ship, whatever the manager's home department. | From instruction to audited work: Hands-on execution, certification and independently audited work. | Readiness check: A practical assessment around day 60 of whether they can run a shift. **Keywords.** Leadership development, Enablement, Training design, Adoption across sites, onboarding program, manager development, curriculum design, certification, coaching, cross-functional rotations, safety audits, scaling a program, learning and development, people leadership, up to eight sites, operational readiness, technical enablement, training at scale, onboarding, multi-site rollout ### The D2C operating platform: one source of truth for 12 buildings - Organization: ShipMonk - Years: 2025–2026 - URL: /#case/dtc - Lenses: Product & Platforms, Applied AI & Decision Systems - Evidence for: Product Operations, Product Management (PM), Strategy & Operations (S&O) / Business Operations (BizOps), AI Deployment / Deployment Strategy - Summary: Planning, live execution, quality and reporting for ShipMonk's D2C domain in one platform, built with Product and Data. **Context.** Operators had no single source of truth. There was no internal tool for live dashboards, so they used Metabase and the ShipMonk app, HR ran on a separate platform, and nothing existed for shift planning or tracking. About half the metrics weren't reported, including subprocess cycle time. **What happened.** I started with the shift planner and made it the standard in every building. Then, with Product and Data, I rebuilt the planning math, built subprocess cycle time up into bottom-up labor planning, and brought live execution, quality alerts and photo confirmation into one platform. Operations leaders used the same cycle times for continuous improvement. It replaced four to five separate sheets and calculators, and the operating model moved beyond its pilot. **What I led.** With Product and Data: business logic, data lineage, definitions, requirements, operator workflows and adoption across 12 buildings, with the AI-assisted operations layer on top. **What lasted.** The operating model moved beyond its pilot and was centralized, with one platform for operational tooling being adopted across the 12 buildings and the remaining domains as extensions. Every building moved from its own calculator to one shift planner. **Headline figures.** 12 (buildings moved from per-building calculators to one shift planner); 4–5 (separate sheets and calculators replaced); 70% faster (operator decisions in the first AI-assisted operations deployment); 32 (data marts behind planning and reporting) **Results.** - Platform: 4–5 separate working tools and sheets replaced; 12 buildings on one shift planner; 32 data marts behind it; Centralized one platform for operational tooling, being adopted across the 12 buildings, with the remaining domains as extensions - Outcomes: 70% faster operator decisions in the first AI-assisted operations deployment; More accurate headcount planning **What was missing.** One source of truth: Every building kept its own numbers, and different teams read the same events differently. | Live dashboards: No internal tool for them. Operators worked from Metabase and the ShipMonk app. | Shift planning and tracking: No platform at all. HR sat on a separate system. | Metrics: About half weren't reported, including subprocess cycle time. **How it was built.** Subprocess cycle time, bottom up: Cycle time for each subprocess from system events, rolled all the way up into labor planning. Operations leaders also got it for continuous improvement, to see how their own process was doing. | Three levels of planning: A labor plan from the demand forecast, then next-day and same-day shift plans for every building, with a staffing manager turning them into people on the floor. | Live signals to floor actions: Execution and SLA signals during the shift, with quality alerts, photo confirmation and audit triggers routed to the floor. | Data and intelligence: Sigma and Snowflake with 32 data marts. On top, AI-assisted operations, including a retrieval assistant that follows document precedence and user access, and refuses when the evidence isn't good enough. **Keywords.** Product ownership, Operating platform, Data modeling, Requirements and rollout, AI-assisted operations, Adoption at scale, product management, product operations, internal tools, platform, single source of truth, data lineage, business logic, requirements, roadmap, prioritization, Sigma, Snowflake, data marts, SQL, analytics, KPI design, metrics, workforce management, capacity planning, demand forecasting, shift planning, staffing, staffing manager, labor planning, real-time telemetry, quality alerts, retrieval-augmented generation (RAG), AI assistant, access control, change management, rollout, training, cross-functional with Product and Data, multi-site deployment, 12 buildings, direct-to-consumer, operator decisions 70% faster, workforce planning, launch readiness, customer and operator feedback into product decisions, issue triage, lightweight automation, internal AI tooling, workflow automation, AI-native operations, baselines and evaluation criteria, implementation, adoption and measurable value, enterprise rollout ### ShipMonk D2C ownership: a $90–95M business, process and product - Organization: ShipMonk - Years: 2025–2026 - URL: /#case/cadence - Lenses: Strategy, Leadership & Transformation - Evidence for: Chief of Staff (CoS) / Founder's Office, Strategy & Operations (S&O) / Business Operations (BizOps), Special Projects / Strategic Initiatives - Summary: Owning the direct-to-consumer domain across 12 buildings in a role that had not existed: planning, the processes, the reviews, launches and merchant quality. **Context.** ShipMonk's D2C business ran across 12 buildings at about $90–95M in annual revenue, and no role like mine existed. Shift planning didn't exist, labor planning wasn't built from demand in any structured way, only a sheet, and documentation didn't match newer workflows like photo confirmation and batching. **My role.** D2C domain owner, with ownership spanning process, product, technical and C-suite reporting, as well as future solutions and designs. **What I led.** The D2C domain end to end, including the full product: internal tooling, the order management side of the ShipMonk app, and every KPI and metric. Planning, launch readiness, SOP and RCA mechanisms, merchant escalations, and the WBR, MBR and QBR I designed, ran and presented to the CEO, CFO, CTO, CSCO and CPMO. **What lasted.** One set of KPI definitions from the floor to the C-suite, documented processes that matched execution, managers trained to read the system, and a planning and reporting model that was centralized, with the remaining domains as extensions. **Headline figures.** 12 buildings (about $90–95M in annual revenue); 60% → 85% (Platinum and Gold merchant goal attainment); 92% → 96% (packaging quality against a 95% contract); $100K (paid photo confirmation across two quarters) **Results.** - Merchants and quality: 60% → 85% Platinum and Gold merchant goal attainment; 92% → 96% packaging quality, against a 95% contract; $100K paid photo confirmation across two quarters, three merchants - Packaging study: 197 before and after comparisons: 137 improved, 54 regressed, 6 unchanged; 6.41 s mean time saved per compared observation, about 314 hours in total; 117,938 orders in a carton-substitution study - Scope and decisions: 12 buildings, about $90–95M in annual revenue; 3 planned facility designs changed through the Aisle Tax; Committee member of the technology transformation committee, driving AI adoption **What needed to change.** Documentation: Newer workflows like photo confirmation and batching weren't in the operating standards, and quality knowledge sat with the Inventory team instead of in D2C training. | Measurement: Activities had no process-level measures, and different teams read the same system events differently. | Planning: Shift planning didn't exist, and labor planning wasn't built from demand, only a sheet. Nothing connected the plan to the floor or to the weekly review. **How the processes were rebuilt.** Process by process: Every D2C process: physical steps, system interactions, what each scan meant, decision rules, standard work, handoffs and documentation. Some needed correcting, some standardizing, some redesigning. | Measurement from the bottom up: Subprocess cycle times from system events and timestamps, rolled up into productivity, capacity and labor planning, with KPI definitions for quality, shift execution and merchant performance. | Business logic: The labor and demand calculations, the rules behind each scan and decision, and the escalation paths, written down and agreed with every department. | Data lineage: Which source is the authority for each number, so the floor, finance and the C-suite read the same definitions. | Operator tools: The shift planner, live production views and quality alerts that put the measures in front of the people running the shift. | Packaging: Recommended versus actual box choice, a rebuilt box suite, station slotting, replenishment and visual standards, set per merchant and product category, top revenue bands first. | Rollout: A three-day training for leadership, managers and analysts. Managers trained their teams, and Learning and Development took over recertification. **Keywords.** Domain ownership, Executive reporting, Operating cadence, KPI design, Business reviews (WBR/MBR/QBR), business owner, $90–95M annual-revenue domain, general management, WBR, MBR and QBR, C-suite communication, operating rhythm, operating model, SOPs, root cause analysis, escalation management, launch readiness, customer experience, merchant escalations, cross-functional leadership, technology transformation committee, AI adoption, documentation, training program, process re-engineering, metric definitions, strategic planning, process design, planning cadence, goal-setting and OKRs, metrics cadence, executive decision support, cross-functional operating mechanisms, business operations (BizOps), strategy and operations (S&O), B2B customers (merchants), post-sales operations, customer escalations, headcount and capacity planning ### The Aisle Tax - Organization: ShipMonk - Years: 2025–2026 - URL: /#case/aisle - Lenses: Strategy, Leadership & Transformation - Evidence for: Strategy & Operations (S&O) / Business Operations (BizOps), Chief of Staff (CoS) / Founder's Office, Special Projects / Strategic Initiatives - Summary: A building launch in my D2C domain turned into evidence that changed an approved facility design at three planned sites. **Context.** Site 1 was a building launch in my D2C domain, and I was there to make sure the process could run and people were trained. A manager told me pickers were hitting constraints in the narrow aisles and work was getting slower. The narrow-aisle design had already been approved to add storage and was moving through the facility program. **What happened.** Active construction was underway at the first of three planned buildings. I built the evidence from production telemetry, matched workloads across the two buildings so the work was comparable, and ran same-picker trials across shelf setups. I took it to my manager, the Chief Supply Chain Officer, who brought in the CEO. The next step was live model experiments, and their results became a design change for the three planned sites, including a retrofit of the one under active construction. **What I led.** Data collection, data modeling, telemetry and process modeling, the picker experiments, executive reporting and the updated design, alongside my domain's launch readiness at Site 1. I had no capital-approval authority. **What lasted.** Designs changed at three planned sites: Site 2 under active construction, and Sites 3 and 4 before construction began. **Headline figures.** 3 (planned facility designs changed, one under active construction); CEO (decision after live model experiments); 756K (pick events behind the evidence); Launch (in my domain, where it started) **Results.** - Decisions: 3 planned facility designs changed; Site 2 retrofit while under active construction; Sites 3 and 4 two sites changed before construction - Telemetry (observational): 756,164 pick events, 685,115 valid; 25 s / 75 s typical pick, median and 90th percentile; 110 s / 257 s narrow-cart group of 2,734 picks, median and 90th percentile; +83 s weighted difference per pick across 835 matched products **How the evidence was built.** Telemetry first: Production telemetry modeled as far as the data alone could go. | Make the work comparable: The buildings handled different merchants and products, so I built matched workloads in SQL: size, weight, handling, reach, location, picking sequence and cart. | Same-picker trials: Weekend test windows, moving the same pickers across locations and shelf setups. Several narrow layouts were tried. Some were fast for small items, but none held up across the expected product mix. | What the storage costs to run: Extra walking, repositioning, tight cart moves, waiting and the slow tail. The question became what the extra storage costs to run every day. **Taking it up.** My manager, the Chief Supply Chain Officer: He brought in the CEO. | Live model experiments: The results became a design change for the three planned sites, including a retrofit of the one under active construction. **Three planned sites.** Site 2: Under active construction. The design changed and the building was retrofitted. | Sites 3 and 4: Still in design. The layout changed before construction began. **Keywords.** Experiment design, Executive influence, Facility economics, Telemetry analysis, Business case, CEO decision, changing an approved decision, controlled trials, matched comparison, SQL, causal caveats, cost modeling, capital expenditure, design change, facility design, process modeling, data modeling, evidence-based decision making, launch readiness, executive reporting, 756K pick events, business case, ROI, cost-to-serve, quantitative analysis, decision memo, strategic analysis, labor economics ### Photo confirmation: a paid merchant service at ShipMonk - Organization: ShipMonk - Years: 2025–2026 - URL: /#case/pack - Lenses: Product & Platforms, Applied AI & Decision Systems - Evidence for: Product Management (PM), Product Operations, GTM Strategy & Operations - Summary: Packaging and kitting escalations turned into photo confirmation on every pack station, switched on per SKU for merchants and internally for packaging issues, and launched as a paid merchant service. **Context.** Packaging quality and kitting escalations kept coming from merchants, and there was no visual record of what left the building. **What happened.** Photo confirmation came out of packaging and kitting escalations. I set it up on every pack station so any workflow could be switched on: per SKU for merchants, and internally for packaging and workflow issues. I offered it first to one heavy-kitting merchant and launched it as a paid service merchants could use as evidence. **What I led.** Product and operations for photo confirmation as D2C owner: the service model (which SKUs and stations, what the camera captures, how merchants see each photo) and the paid launch. **What lasted.** A paid service, $100K across two quarters with three merchants. The vision questions it raised continue in the AI Lab. **Headline figures.** $100K (paid photo confirmation across two quarters); 3 (merchants on the paid service); Every station (set up so it could be switched on per SKU) **Results.** - Service: $100K across two quarters; 3 merchants on the paid service; Paid value-added service **Photo confirmation.** Where it came from: Escalations on packaging quality and kitting. | First merchant: Offered to one client with heavy kitting, showing how it caught problems before they shipped. | How it ran: Set up on every pack station. Merchants switched it on per SKU for at-risk items, and we switched it on internally for packaging and workflow issues. Merchants could view each saved photo, which gave them evidence when escalations came from their supply chain, transport or their own customers. | Rollout: Three merchants, as a paid service. **Keywords.** Service launch, Visual evidence, Quality, Customer experience, Monetization, value-added service, paid service, customer evidence, escalation reduction, camera capture, product launch, pilot customer, kitting, quality assurance, human-in-the-loop, vision research follow-on in the AI Lab, $100K in two quarters, every pack station, per-SKU switch, commercialization, packaging and monetization of a service, customer pilot, B2B ### Research partnerships - Organization: A.R.I.E.S. (technology startup) - Years: 2012–2019 - URL: /#research - Summary: Through the startup, four Centers of Excellence and a LabVIEW academy (250+ certifications) behind the NI market, and the research platforms built on them: a flight-tested airship (papers 2014–2025), a Massive MIMO platform, an assistive wheelchair (NI Asia award) and a robotic bowling arm. **Keywords.** Centers of Excellence, academic partnerships, research platforms, Massive MIMO, 5G research, airship, remote sensing, thermal imaging, assistive technology, multimodal input, robotics, mechatronics, FPGA, myRIO, LabVIEW academy, 250+ certifications ## AI Lab (current work) - /#labhome/engine: How I work, the design of a research engine being built: agents run experiments, a deterministic kernel holds the record, a model that did not do the work checks results (from a separate provider where possible), and failures go to Detention for a root-cause analysis; a domain advisor agent is being added - /#labhome/vision: Vision, in active experiments - /#labhome/detention: Agent Detention, in active experiments - /#labhome/world: World model, in planning ## Notes for agents - Modeled figures are labeled as modeled (for example the Prime Day exposure of about $12M a day). - Some client, institution and site names are withheld for confidentiality. Ask in an interview for detail. - Structured data: /profile.json and /resume/resume_data.json.