Live — Web + Apple App Store
Daily word game with a persistent generative character that reacts to real play. Deterministic code owns the puzzle, score, and limits; the model produces bounded character behavior from structured game context.
- Server-owned rules prevent the generative layer from ever touching game truth — score, answer, and limits are unreachable from the model.
- Character stays recognizable across guesses, hints, difficulty modes, wins, losses, fallbacks, and model changes.
- Automated regression checks gate every release, including character-behavior changes.
- Three daily difficulty modes, themed games, contextual hints, and shareable challenges.
- Capacitor ships the public iOS app from the same product core as the web game — one implementation, two surfaces.
reacttypescriptgeminicapacitorsupabasevercel
Live — Web + Apple App Store · concept to App Store in 14 days
Multimodal social game that turns social photo-grid screenshots into a playable friend quiz. The model proposes the answer key; code bounds what a wrong one can reach.
- The model's guess ships as the answer key by default, so machine-checkable invariants bound it: no answers in the player payload, reveal only after a valid pick, an immutable seal, and scores only for players who actually played.
- Privacy-preserving by construction: creator-supplied screenshots, independent per-photo verdicts, no facial recognition, no Meta API access, no staff photo review.
- Early soft-launch generation cost measured at roughly 6–7¢ per completed quiz.
- Scoped to objective visual memory over personality judgment, so the model never invents claims about a person.
- Answer-specific reactions, named friends, and timed group reveals drive the social loop.
multimodal aicapacitorserverlessweb + ios
Live — Web
Competitive AI game built so the model can't be gamed — structurally, not with prompt hardening. A player actively trying to manipulate an LLM, on a paid API, in front of a family audience.
- The model never emits the score. It returns taste only — five bounded axes plus one in-voice line — and the favor delta is derived server-side by a pure function the model cannot see or reach.
- Denial-of-wallet bounded five ways: 3-turn daily cap, per-IP and per-user rate limits, atomic concurrency slot, cached degrade mode, and a hard spend stop.
- Deterministic red-line pre-filter runs before any model call, so flagged input costs zero tokens.
- Prompt injection handled structurally — the player's reply rides as data being judged, never inside the system instruction.
- Cached degrade mode serves an in-voice reaction with zero model calls under load; the meter never breaks and the character never breaks.
reactgeminivercelsupabase
Open source — Python
Six synthetic executive lenses with competing incentives audit an AI-built operating plan that forecast ~$65.9M in profit. Rebuilt as a transparent, reproducible model when consensus alone failed to catch the arithmetic.
- Reconciled a $3.397M gap between reported and fully loaded base-case expenses that survived six rounds of review.
- 100,000 seeded Monte Carlo trials stress-testing sell-through, revenue execution, expense variance, weather, sponsorship, and delays.
- Median reconstructed profit $1.22M; probability of reaching at least $5M: 10.5%.
- Invariant tests block incompatible expense definitions from quietly surviving consensus.
- Deterministic outputs, tests, evidence ledger, and limitations published publicly and labeled as reconstruction.
pythonmonte carloregressioninvariant tests