Hi, I'm Chris.
Over a decade of production software built with the customer in the room, for humanitarian programs, federal agencies, regulated fintech, and my own shipped products.
I scope problems that arrive unspecified, build the system end to end, and instrument it so the customer can see whether it worked. Most of what is linked here is public and still online.
Open to forward deployed roles
Talking to teams building applied AI, technical field organizations, and anyone whose hardest deployments happen inside someone else's systems.
The four phases below are how I work. Each one is written against PlateQuest, a shipped iOS product I own end to end, so none of it is hypothetical.
01 Discover
Four numbers agreed up front: 95% top-one precision at confidence 0.9 or above, 30% coverage, sub-100ms inference on the Neural Engine, under 10MB bundled. Precision leads because a confident wrong answer costs a player a first find, and paying for a cloud call is the cheaper mistake.
02 Prototype
Two standalone spikes ran before any architecture was committed. Nearest-anchor retrieval on Vision feature prints reached 52.2% top-one with no training step at all, against a production softmax baseline of 40%. Text recognition read the state banner on 84.1% of real captures for roughly 40ms.
03 Productionize
The on-device classifier answers what clears the gate, anything below it falls through to a cloud vision model, and what the cloud cannot settle lands in a human review queue. New architectures run in shadow mode behind a remote flag, writing telemetry but touching no user-visible decision, before they take the default path.
04 Instrument
Telemetry records every cloud prediction, auto-approval, and admin correction, and an internal dashboard reads precision, coverage, and a confusion matrix off live traffic. The retrieval successor scored 64.6% against its 70% gate on a 48-capture holdout and did not ship.
The full evaluation is published, including the run that failed: 6 registered policies against the 70% gate, the precision and coverage sweep, the 17 misses sorted by cause, and per-class accuracy with the sample sizes left in.
Read the eval resultsEight products for a USAID program, UNHCR, and civic, trade, and agriculture audiences. Six are still online, several rebuilt years later off their legacy stacks.
A market-oriented dashboard for understanding export behavior and shipment movement through concise comparative views.
The USAID Financing Ghanaian Agriculture Project's investment mapping system, rebuilt as a static atlas: 53 northern districts of observed crop production, archived farm geography, road networks, and soil classification on one MapLibre canvas.
A county-level explorer for USDA QuickStats covering corn and soybean yield, production, and acreage, pairing a national choropleth with a distribution histogram and a full-history timeline that scrubs the map.
A split-screen Mapbox GL map that sets two decades of American Community Survey median household income side by side, with a draggable divider for reading how income shifted tract by tract.
A Mapbox and D3 explorer, built with UNHCR for The Opportunity Project, mapping where resettled refugees were placed across U.S. cities over a decade and overlaying the charitable resources available nearby.
A state-by-state timeline that sets daily case and death curves against the opening and closing orders that bracketed them, so each policy decision can be read against what actually followed it.
A constituency-level product that translates fragmented political signals into a clear map, filter system, and representative view.
A weather analysis prototype combining time, geography, and storm intensity in one view to support rapid pattern recognition.
A regulated lender, a global data-collection platform, a bespoke visualization firm, and a studio of my own. In each one the scope arrived from the people who would use the result.
Full history on LinkedInSr. Principal Engineer · Regulated consumer fintech
Front-end architecture for a consumer credit product, across web and React Native. Principal from 2022, senior principal from 2026.
Founder, product studio · Data products and AI apps
Own studio shipping end to end: TradeNet, the USAID FinGAP atlas, the USDA NASS explorer, and two App Store products with live model telemetry.
Advisor · Nonprofit
Technical advising on architecture and delivery for a small nonprofit engineering effort.
Staff Frontend Engineer · Global data collection
Architected and implemented an organization-wide design system and component library in React, on Material UI and Base Web, plus the tooling and deployment process around it.
Data Visualization Engineer · Mission-driven client work
Bespoke visualization delivered directly to clients: UNHCR resettlement mapping for The Opportunity Project, census income comparison, and the UK Political Atlas, built alongside award-winning information designers.
Visualization internals, design systems, and testing practice, written the way a deployment write-up is written: what was built, what it cost, and where it would break next.
Visit Medium
Jun 8, 2026
A 3D view of global trade that scores every relationship against what a gravity model predicts — and flags who should trade more.
Jun 2, 2026
Engineering custom, award-winning visualizations with d3.js and React/Angular
Nov 12, 2025
A React Native team’s journey to AI-assisted testing, visual truth, and cross-functional confidence.
Mar 27, 2025
Using Percy.io to automate screenshot testing across Storybook and a matrix of browsers and screen sizes
Reservation workflow, VR behavioral analytics, and hierarchical visualization, each prosecuted through examination to grant.
US 10,860,954
A mobile reservation workflow that uses OCR to identify a conference room, surface its live schedule, and book the room or suggest nearby alternatives when it is unavailable.
US 10,747,760
A visualization and analytics system for VR experiences that captures user movement and position data, then turns those signals into heat maps and reviewable behavioral patterns.
US 10,482,130
A method for rendering hierarchical data as three-dimensional tree diagrams by positioning nodes and edges with structural constraints to improve readability and analytical exploration.
Send the role, or send the problem that keeps stalling. I am glad to walk through the PlateQuest evaluation in detail, including the parts that did not work.