I build the parts of a product that have to work under pressure.

Onsale

A ticket onsale engine for when thousands of people want the same seats

A virtual waiting room, all-or-nothing seat holds and idempotent checkout, rebuilt in the open from the problems I worked on in production ticketing.

• Redis + Lua waiting room admits buyers in arrival order at a fixed rate, so the seat map and payments only see as many shoppers as they can serve.
• Picking four seats holds all four or none in one atomic step; holds expire on their own.
• Checkout is idempotent: a double-click or dropped connection replays the first result instead of charging twice. A Postgres constraint makes selling a seat twice impossible.

696/696

seats sold in ~20 s to 2,000 simulated buyers

0

seats sold twice, audited in Postgres after every run

0

unexpected errors in 24,251 requests

TypeScript, Fastify, Redis + Lua, PostgreSQL, React, k6, Prometheus, Grafana

VoiceBrief

Meeting notes you can check, with every item linked to where it was said

Record a meeting and get decisions, action items and open questions. Ask questions across all your meetings and get answers with a citation for every claim.

• Every note carries evidence: the model cites transcript lines, and a grounding pass drops anything that cites lines that don't exist or names an owner who never spoke.
• Hybrid retrieval (keyword + vector, fused by rank) answers questions only from cited excerpts, or says plainly that something wasn't discussed.
• Transcripts are treated as data, not instructions, and an evaluation harness scores the pipeline against hand-labelled meetings.

Next.js 16, TypeScript, Whisper, LLM structured output, Embeddings / RAG, pgvector

MediaVault

A shared media library for creative teams, built as a real multi-tenant SaaS

Teams upload pictures, video and PDFs straight to storage, find them by title or tag, collect selects and send clients expiring links.

• Each team's data is isolated by Postgres row-level security, so a careless query returns nothing rather than another team's files.
• Uploads go straight from the browser to object storage with signed policies; a background worker checks each file really is what it claims and renders thumbnails.
• One role and permission table is shared by the API and the UI, and a test spells out every role and action pair.

TypeScript, Node.js, PostgreSQL RLS, S3 / MinIO, Background jobs, React

Groundwork

Survey and CAD data in the browser, measured to the millimetre

Open LandXML terrain and DXF site plans, then measure distances, areas, long sections, stockpile volumes and cut/fill against a design. Everything runs on the visitor's machine.

• Coordinates stay in float64 and only move to a local origin at the GPU, so readings at E 411 020.373 are exact instead of snapping to 6 cm.
• Volumes against a plane are exact (triangles are clipped where they cross it), and cut/fill between two surfaces is checked against the exact result in tests.
• A grid index answers a million elevation lookups in 0.2 s; parsing and volumes run in a Web Worker so the 3D view never stalls.

40

tests checked against answers known in advance

0.12 s

cut/fill between 29,000 triangles of ground and design

0

bytes uploaded: files never leave the browser

TypeScript, three.js, Web Workers, React, LandXML, DXF, Vitest

ShelfSync

A point of sale that keeps selling offline and never loses or doubles a sale

Two tills and a back office. Switch a till's network off, sell, switch it back on, or watch both tills sell the last mug while offline and see how the conflict is resolved.

• Tills save every sale to an outbox in IndexedDB first and sync what happened ("sold 1"), not stock levels, so sales from several tills never overwrite each other.
• Sync is idempotent and ordered: replays get the server's first answer, and a per-store advisory lock keeps sequence numbers gap-free so no till misses a change.
• A sale that already happened is never refused. If it takes stock below zero it is kept and flagged for a recount, and a recount keeps other tills' sales instead of erasing them.

200

random chaos runs (offline tills, lost and duplicated requests) that all converge

0

sales lost or counted twice across 24 concurrent syncs

22

tests against a real Postgres, run in CI

React, TypeScript, IndexedDB, Express, PostgreSQL, Netlify Functions, Vitest