Bot Busters

Bot Busters drops you into a group chat with other players, one to three of whom are LLM-driven bots trying to pass as human. After a timed chat round, everyone votes on who they think the bot was. I joined a month into the project and spent ~4.5 months as the backend engineer building the parts that make the game fair and worth playing again.

What I owned

• The match / game-state machine (match.ts): an in-memory class, one instance per active room, driving stage transitions (chat → voting → results) off wall-clock elapsed time. It includes the server-authoritative anti-cheat decision I’m proudest of: the isBot flag is stripped from every outbound payload until the reveal, so no client can peek at a secret it can inspect over the wire.
• calculatePoints() — the whole reveal scored in one pass: bots busted, humans wrongly busted, humans fooled.
• An achievements system as a small rules engine — eight self-contained achievement records ({name, description, calculate}) evaluated against one shared context object. Legible enough to walk through end to end, shipped across six PRs.
• Wallet + anonymous auth — anonymous play as a first-class mode, not a fallback, sitting next to wallet-verified identity that gates which achievements count.
• DB access-layer extraction and a match-history join table (user_match) that replaced deserialising every match’s JSON blob just to answer “how many matches has this user played?”

What I did not build — said plainly

The thing most people find interesting about this project is the bot-AI itself, and that’s not my code. The personality-driven prompting, the randomized read-delay and typing-time pacing, and the seven-model / four-provider Bedrock roster were written by a teammate (Ramiro), in files I have zero commits on. My contribution was the system those bots plug into — the match, scoring, and achievements that decide whether a bot actually fooled anyone. “I built the game the bots had to survive” is the honest way to put it.

A real tradeoff, not a highlight reel

The repo had exactly one test file and one CI workflow that auto-deployed every push to GKE staging with no test or lint gate. That’s a move-fast, small-team setup — and it’s worth saying out loud rather than implying a rigorous test culture existed.