Applied AI engineer. I build LLM systems that go in front of real users — and the evals that tell me whether they actually work.
Available for contract work: fixed-scope builds and monthly retainers for teams in the US, UK, EU and APAC.
Most of what's below I built as the only engineer on it. The farm platform I also run in production myself — nobody else is on call for it.
Working on now
- friendly-memory — long-term memory for LLM companions, and the harness that measures whether it actually remembers. An over-cap write is a hard error that forces consolidation instead of silent truncation. Ships a persona harness that quizzes from ground truth, screens answers for embellishment, and prints its own judge-disagreement rate as the noise floor. 114 tests.
- Chanakya (client work, private) — retrieval over 59,211 Supreme Court of India judgments. 1.3M vectors, 25 GB corpus, precedent as a citation graph, each case framed for counsel, litigant, bench, and defendant. 30 rounds of retrieval eval against probe sets: landmark cases, citation probes, adversarial phrasings, unanswerable questions.
- Nesting Greens (my own farm business) — D2C storefront, WhatsApp ordering bot, and a ~43,000-line operations platform in daily use by a 10-person team, none of whom had used software before. 1M+ eggs sold through it. Containerised on a shared VPS with nightly off-site backups and self-healing timers, extended into ESPHome shed controllers. Nobody else is on call.
- Agent fleet — ~2,500 lines of shell managing a fleet of Claude Code sessions, one per project, each addressable from my phone through its own Telegram bot, forking mid-conversation into isolated git worktrees. Everything above from the last year was built through it.
What I reach for
Python TypeScript React / Next.js Node Flutter PostgreSQL Supabase Neo4j Docker nginx ESPHome
Gemini Live API · Claude · RAG and retrieval pipelines · embeddings and vector search · tool calling · agent orchestration · context and memory design · eval harnesses and LLM-judge scoring · realtime voice (VAD, barge-in, latency tuning)
Earlier
Ran a product consultancy for early-stage web3 teams through 2022–23 — an NFT marketplace built with Verzetsmuseum Amsterdam, a 43% minting-gas cut on a customized ERC-721a, a GMX fork moved from Chainlink to Redstone oracles. The zk and Solidity repos pinned here are from that period.
I take on fixed-scope engagements and monthly retainers, direct or white-labelled through agencies. Working UTC+5:30.
| Retrieval & eval audits | Probe sets on your corpus, a scored baseline, the judge's own disagreement rate as your noise floor, a ranked failure taxonomy, and the harness wired into your CI. |
| Scoping sprints | Architecture, retrieval and eval plan, a prototype on your data, and a fixed-price quote for the build. |
| Conversational commerce | WhatsApp ordering, support and operations automation, end to end. |
| Production RAG, agent and realtime-voice builds | Scoped, built, evaluated and deployed. |
Tell me what you're stuck on and I'll tell you honestly if I'm the right person for it.