An AI mystery game where players interrogate AI suspects to solve crime-scene cases. Live on iOS and Android since June 2026. I own the backend: LLM serving, an agent-driven content pipeline (one session taking designer, writer, and reviewer roles in turn), and payments.

_{GA4, as of Sep 26, 2026}

Constrain agents with code, not prompts. Our authoring agents started approving their own checkpoints. Now checkpoint transitions go through a deterministic gate that re-runs the validator itself instead of trusting the agent's report. → write-up

Decide with production data. A 50/50 production A/B test rejected a model swap that looked right on paper (p95 1.5 s → 4.5 s, +57% cost per call). The same test showed our cross-model fallback recovering 82 of 83 paid requests during a provider outage.

Make money paths exactly-once. Idempotency keys plus Firestore transactions: zero double charges across 60 runs of 20 concurrent same-user S-coin spend requests (test env). Requests that can never succeed are quarantined instead of retried forever. → write-up

Tie cost to the right variable. Designed the notification inbox so cost grows with active users, not with subscriber count. → write-up

Co-founded Tecketing in 2024 and ran it as CEO until 2026, then moved to Tech Lead. 6 startup competition awards · KOCCA commercialization program · GSIA 2026 Seattle global program · 61 straight weekly sprints

• Game scenarios: from outsourcing to our own pipeline — moving scenario authoring from outsourcing to a gated agent pipeline
• In-app payments, Part 3: the problems that start after verification succeeds
• FastAPI and cloud native, Part 1 — why LLM calls moved to the server

• [스타트업/기술] Sleuth 프로토타입에서 프로덕션까지 | 전체 포스팅 모음
• [스타트업/기술] 에이전트의 Self-approval 문제 | LLM 저작 파이프라인을 결정론적 게이트로 막다
• [42 글로벌프로그램] 다시 미국에 오다 - (3편/완결)
• 추리 게임 시나리오, 외주에서 자체 파이프라인으로
• [42 글로벌프로그램] 다시 미국에 오다 - (2편)