Work

HeySomm is a live wine platform I designed and run end to end. Scan a label, and a vision model, retrieval over pgvector, and an eight-axis match engine turn it into a recommendation, streamed back over SSE. The part I'd defend hardest isn't the model work, it's the discipline around it: six scan stages run on versioned, A/B-routable pipelines, and nothing ships until it has been replayed through the real inference path against recorded scans. It runs on a Hono API and a native iOS app, with multi-provider inference — Anthropic, OpenAI, Cohere — behind circuit breakers, deployed blue-green on Fly. It builds a palate profile for each drinker and scores every bottle on quality against price, so the recommendation is personal, not a crowd average. An in-house LLM-ops layer governs six scan stages behind a replay-before-publish gate.

A multi-tenant legal-AI retrieval platform I'm building. A vector store was the obvious choice, and for most corpora it would be the right one. This is regulated legal text, where every answer has to cite something a lawyer can check — so it runs on lexical retrieval with a cross-encoder reranker and typed citation verification: citations are classified — legislation, case law, court rules — and checked, not just generated. Four LLM providers route through one abstraction with per-slot fallback. The knowledge base is fed by a corpus pipeline of twenty-one scraper adapters — nine Australian jurisdictions and twelve regulators — producing a 220 MB standardised corpus. Client and tenants stay unnamed by agreement. Interpretable beat fashionable.

A multi-tenant CMS platform for a governance consultancy — several branded sites from one Next.js and Sanity codebase, eighteen months of single-author production history. This site runs on the same engine.

Enterprise order orchestration: about ten storefronts into a Cargowise 3PL and a Dynamics NAV ERP, over REST, XML and SOAP, with multi-tenant SaaS patterns. Senior engineer on the team; architecture shown, not linked.

Personalisation storefronts for Victoria Bitter, Carlton Draught, Great Northern and Carlton Dry, plus Nestlé campaigns, notably Milo — an embedded design editor producing print-ready PDFs. Delivered with a team some time ago; the storefronts are still live.

arbitron, a Polymarket CLOB arbitrage bot, and MemeScreener, an XState actor system across Sui, Solana and EVM. Their math, my systems — live on real capital.