Paste a research paper’s link, DOI or title and read it as a clean web page — then discuss it in your own language, out of 70+. Built from prototype to the live site.
WORK
Everything I’ve built
Shipped products first, then the platform work underneath them, then prototypes and research. Where something is a prototype, or was operated rather than built by me, it says so.
Shipped & live
Tracks how brands get mentioned and cited across seven AI answer engines. I specified the product and its metric model, owned the unit economics, and directed AI coding agents to build it.
Checks whether a brand shows up when people ask AI assistants about its market. Runs as a command-line tool and an MCP server, graded against human labels in a gate that fails the build on regression.
Platform & data
The data layer under the GEO metrics: captures answers and citations from the major AI assistants three ways — official APIs, paid scraper vendors, and custom clients — across three deployment targets.
Financial report data extraction
Pulls the key details out of bank research reports automatically. Ten vision-model setups compared on one fixed test set, then the winner run over 14,105 real reports at roughly a tenth of a cent each.
Conversation data reliability pipeline
Turned years of scattered team chat and call transcripts into one clean dataset, cutting unidentified speakers from 3,119 to 964 — with every original message kept beside its AI-cleaned version so any edit can be undone.
Prototypes & client work
Open textbook finder
Finds free, openly-licensed replacements for paid university textbooks — and says “no match” rather than inventing one. A deliberately unanswerable query was correctly rejected across all 15 candidates.
Mobile ID & inventory scanner
A phone camera that reads printed ID and inventory numbers as you point it. The pipeline waits until the frame is steady and actually contains text before spending money on a vision call.
Buy-Canadian product scanner
Point a camera at a product and find out who really owns the brand, with Canadian alternatives grouped by province. Three AI providers benchmarked — then a much faster architecture reverted once it started making mistakes.
Research & infrastructure
Document-understanding research
Tested whether text on a page could be matched the way DNA sequences are, using a bioinformatics alignment tool on vectorised page images. It did not work, and I established why — reported rather than quietly dropped.
A self-hosted agent platform I run and extended: a fully offline WhatsApp voice assistant with no cloud speech, plus scheduled jobs — one captured 6,277 Google AI Overviews overnight. I did not write the platform itself.
An always-on system giving teammates in the US, Netherlands and Kenya their own private browser-reachable desktops — plus a live server migration run through AI coding assistants with no downtime for the team.
