Hands on, not just spec work

A Monte Carlo simulator for gacha banner planning. FastAPI backend, React and Vite frontend.

The Monte Carlo simulation, the probability logic, parameter handling, storage and retrieval, re-run comparison, and the agent layer.

I wrote specs and technical contracts, then directed Claude Code on frontend implementation and a targeted refactor that exposed hardcoded parameters as user adjustable settings. I reviewed and modified the output rather than accepting it as is.

The agent originally interpreted a single simulation result. That solves the wrong problem, since a user can already read a probability output themselves. What they actually need is to know what to do next. So I rebuilt it. The system surfaces deterministic suggestions from the results, and when the user acts on one, the agent re-runs and interprets the delta between runs rather than the snapshot.

Same principle as the enterprise work. Make the output actionable, not merely informative.

Pulls daily OHLC data through a market data API, captures chart context through automated browser interaction, and produces both a deterministic computed signal and an independent AI read.

The browser automation and vision model design, the pipeline logic, AI web search based parameter recommendation, and the market data integration.

Computed and AI reads are displayed side by side, and the interface explicitly flags when they disagree, telling the user to check the source chart. It does not merge them or pick a winner.

The same trust principle as the agentic work, in an unrelated domain. Pipeline timing, capture, analyze, and data steps, is instrumented and displayed, so performance is visible rather than a black box.