Things we’ve shipped.
Work we have audited, specified, built and shipped. In most cases we still run it. More to come as we publish them.
Angkor Now
A self-maintaining city guide for Siem Reap, launched 2026. An LLM pipeline reads the web, extracts structure, de-duplicates and stages everything for one human approval before publishing. Built spec-first; still under our Maintain service: monitored, evaluated and maintained.
Token-Spend Toolkit
One normalised data layer for Anthropic and OpenAI usage, seven tools over it: forecast and budgets, a self-contained Command Centre dashboard, leave-one-out anomaly detection and ranked, dollar-costed actions. Derived cost reconciles to within a millionth of a dollar; 204 passing tests pin every figure.
Index
A rental-search demonstrator for Siem Reap that shows real-estate agencies what neighbourhood intelligence looks like: every listing scored against a stated brief, amenities recorded with tri-state honesty, and a deterministic ingestion pipeline that reads agency sites, extracts structure with zero AI calls and stages everything for one human approval. Configured once; the pipeline keeps it current.
Location Impact Explorer
A geospatial decision-support demonstrator for clean-cooking programmes across Senegal, Mali and Cambodia. A build-time Python pipeline turns GHSL, Kontur and WorldPop rasters into H3 hexagons classified on UN DEGURBA-aligned thresholds; Leaflet and Chart.js render a map, an operations dashboard and a household registry, every figure computed on load from 300 labelled synthetic records. No server, no build step, badge on every screen.
Pendant Rig
Two orphaned AI wearables — a Limitless Pendant whose cloud service is dead and a PLAUD NotePin that deletes recordings into a subscription app — captured entirely offline from one Windows PC. BLE sync lands in append-only journals, a fail-closed codec decodes Opus, and faster-whisper transcribes on a local GPU. The pendant frees flash only after audio is flushed and verified; 900 tests hold the line.
