BTC 5-min Polymarket Bot

Trading Bitcoin's 5-minute prediction markets is normally a four-person job — this project asks whether one engineer can do it.

The market reprices every 300 seconds, so a bot has to read 11 live exchanges and decide in well under a second. Doing that reliably usually needs a data engineer, a backend engineer, an SRE, and a quant. The question here: can one engineer with AI agents, under strict review gates, build and run the whole thing.

A live Rust system on AWS prices a bet and decides in 21 microseconds — measured, not estimated.

Two always-on processes run on one small ARM box. The collector reads 11 exchanges and captured 291 million market events in a single day. The trading process scores a machine-learned model on every order-book change; production tracing puts the full decision pass at a 21 µs median over 506,243 calls. It runs in paper mode — every order takes the real signing and venue path but is rejected at size zero, so nothing is risked.

Nothing polls — every decision fires the instant a market event arrives, moved between processes in about a microsecond.

A venue message flows through lock-free shared memory in ~1 µs and either changes a decision or doesn't. Orders are pre-signed in a pool, so the hot path never waits on cryptography. Every order writes a timestamped audit row for each state change — replayable transition by transition.

13 quality gates block every commit, and there is no flag to skip them.

Each commit passes link checks, schema-drift detectors, supply-chain audits, and an AI code reviewer that can block it — no bypass exists. The suite holds 2,948 passing tests, and 102 soak-test runners check the live data against a 5-minute mirror. An offline simulator replays recorded data and reproduces every live decision at 100% parity, so changes are proven safe before they hit the wire.

2,872 commits and 215,000 lines of Rust in 84 days — solo, with every number measured on the running system.

One engineer shipped this in 84 days with AI agent orchestration as the only toolchain. Coverage held at 99.8% of capture windows over 19 days, and 65 frozen design records explain why each major choice was made, so any future engineer inherits the reasoning — not just the code. The same discipline (spend caps, audit logs, unskippable tests) applies to automation work an order of magnitude simpler.