satinlabs / README.md
Hi, I'm Satin 👋
I came to AI from operations and project coordination in retail and hospitality. I start from how work actually happens, then build systems that take the repetitive parts off people's hands.
- The AI talks; the code decides prices, dates and bookings.
- Every system ships with a test set, and I report the numbers, including changes that made things worse.
- Work starts on the cheapest model that does it well and moves up only when needed.
Results
| Project | Measure | Result |
|---|
| Model Arena | Clients on one API | Web + iOS + Android |
| Model Arena | Cloud vs own server speed | ~400 vs 5 tok/s |
| LINE AI Receptionist | Test conversations passed | 9 / 9 |
| LINE AI Receptionist | Prompt-injection booking attempt | Blocked |
| Agent Lab | Independent judge score | 7.4 → 8.4 |
| Agent Lab | Steps finished by free models | 40% → 47% |
Pinned
Ask once, watch several AI models stream answers side by side, then vote blind. Leaderboard split by Thai and English. Cloud models race a free Llama on my own server.
Native mobile app for Model Arena. Live streaming answers, blind voting and leaderboard; shares one API and database with the web app.
JavaScriptReact Native · Expo
AI receptionist for Thai shops. Answers from the shop's own facts, books and cancels under rules enforced in code, works in LINE and on the web. Auto-deploys on every push.
Cost-aware multi-agent system. Planner, workers and reviewer; each step escalates from a free local model to Groq to Claude only when the reviewer rejects it. Scored on a fixed eval set.
Pythonjudge 7.4 → 8.4
Build log
Less More
- arena-mobileiOS/Android app with live streaming via expo/fetch
- model-arenaDocker + auto-deploy to arena.satinlabs.dev
- model-arenaParallel SSE streaming, blind voting, Thai/English leaderboard
- line-receptionistAdd LINE webhook with signature check
- line-receptionistAdd GitHub Actions auto-deploy with health check
- line-receptionistChat eval: 9 cases incl. prompt injection; fix reply language
- line-receptionistBooking tools with SQLite; shop rules in code; calendar fix
- agent-labv3: ready-to-use outputs (judge 7.4 → 8.4)
- agent-labEval harness: 5 cases, two judges, baseline
- infraServer hardening, Docker, Caddy HTTPS for satinlabs.dev