Nov 2026 – Now
AI Engineer
A Bangkok AI lab building character memory and AI companion products. I work on speed, cost and long-term memory for its production chat service.
- Built a parallel version of the production chat service behind percentage and per-user switches, so new behaviour could be tested on real users without risk. It became the base for the version players use now.
- Cut model calls per message from 4 to 2 and response time from 6.5 s to 4.4 s with reply style unchanged, and raised prompt-cache use from 61% to 70%, cutting cost per message by 14%.
- Redesigned long-term memory so no turn or write is dropped, raising recall from about 20% to 34% and making background writes 39% faster.
- Built a replay test on 20 real conversations with blind LLM grading to compare 8 memory setups for under $5 a run, and used it to reject six proposed changes, one of which needed a database migration.
- Shipped an AI auto-reply system for the company's Facebook Page inboxes, including BuyMeChocolate, in one day from prototype to production: 5.1 s median reply, up to 10 Pages per machine, each message handled exactly once, and sending guarded by allow-lists, hourly limits, human takeover and an emergency stop.
TypeScript · Node.js · LLM evaluation · Prompt caching · SQLite · Claude · OpenRouter · Ollama


