Snowball — an autonomous trading agent
A cash-only trading agent that learns from its own results. I built the closed learning loop: it logs every decision and outcome, then adjusts its own confidence from what actually worked. The interesting part was the debugging — the desk once graded a batch of trades against a frozen after-hours tape and read a whole "bearish market" story into what was really a measurement bug. Catching the system's own plausible-but-wrong conclusion, and making it fail silent instead of learning from noise, is most of the real work.
Python · Bayesian confidence weighting · paper-trading soak tests · a gated deploy pipeline with auto-rollback.
financial-os — a personal-finance engine that refuses to guess
A transaction-categorization and cash-flow system built on one rule: when it isn't sure, it holds instead of guessing. It splits a gas-station charge into fuel vs. snacks only when a receipt proves the split, and it corrected its own over-matches with a per-row audit trail. When a 60-day forecast came out badly distorted, the fix wasn't the forecast — it was treating the wrong number as a map of what the system wasn't yet observing cleanly.
SQL ledger · receipts-as-ground-truth matching · cadence-based recurring detection · human-in-the-loop confirmation.
Downloads Sweep Agent — an agent that treats its own inputs as untrusted
A file-handling agent that reads, summarizes, and routes downloads for a human to confirm — never auto-deletes, always reversible. I designed it to treat the contents of the files it reads as untrusted data, not instructions, because a file is a claim about itself, not the truth. An early version was an always-on watcher; a design review overturned it for something simpler and safer built for how the task actually happens.
Ports-and-adapters design · dry-run default · reversible staged moves with an undo manifest · indirect-prompt-injection-aware.