AI Model Testing Arena
Runs several LLMs over the same 50+ case test set, validates structured output with Pydantic, scores accuracy, latency and cost, and includes prompt-injection cases. Every run logged to Postgres.
I ship web apps with AI agents in the loop, and I build the agents. React, Node.js and cloud on one side; agentic systems, MCP servers and tool use on the other.

AI in the loop
Projects
Five local builds. Source and demos go up here as each one goes public.
Runs several LLMs over the same 50+ case test set, validates structured output with Pydantic, scores accuracy, latency and cost, and includes prompt-injection cases. Every run logged to Postgres.
Campus RAG with hybrid retrieval (pgvector and full-text), Cohere reranking, role-based document access, and cited answers that abstain when unsure. Ragas eval on 57 questions.
Crawls a site with Playwright, runs axe-core on every page, then asks a vision model the questions rules cannot answer, like whether the alt text is useful. Fails the build on regressions.
Keeps live transcription in sync with slide changes, so a student can ask what the lecturer meant two minutes ago and get an answer grounded in both. Notes and a quiz afterwards.
Hourly demand model on 17k Capital Bikeshare records with a time-aware split. RandomForest cuts MAE from 113 to 47 against a same-hour-yesterday baseline. Predict and retrain endpoints, drift monitoring.
One live site, five earlier builds from 2022 to 2024.
Live now
This site. Prerendered Preact, 25 kB of JavaScript, Lighthouse 100 on accessibility, best practices and SEO, built with coding agents in the loop.
Preact, Vite, Claude Code
Lighthouse
How it is built
Prerendered Preact, 25 kB of JavaScript, self-hosted fonts, no tracking.
Earlier builds
React, MongoDB, JWT
React, LocalStorage
React, Redux, Express, MongoDB
React, Express, PostgreSQL, WebSockets
TypeScript, Express, Docker, AWS
Experience
Websites and web apps shipped with coding agents in the loop, and agentic systems and MCP servers built on top of them.
Contract role at Harbour Net. Technical support and ServiceNow administration against SLA targets.
2017 to 2023: customer-facing roles at Woolworths, Highfield's Huxley and Mandala Nepalese Restaurant.
How I work
I build websites and web apps. Most of that work now runs through AI: agents that scaffold, test and ship alongside me, and agentic systems I build. My usual day is where I try new ideas and keep the ones that hold up.
Building for the web since 2016, shipping professionally since 2022. What stuck is the habit of taking a thing from the first sketch to the deploy log, and answering for it afterwards.