I make machines useful. Occasionally clever.

Built to turn learning activity into an actionable view of learner performance — not just predict a result, but surface who might need support and why.

I inherited an early prediction-engine scaffold and was the main contributor during its functional build and design phase: reworking the prediction contract and designing and implementing the rule-based logic behind it.

Where it gets interesting is what the system refuses to pretend. It defaults to rules-only, with ML gated behind explicit data, validation and monitoring criteria. And confidence is data-completeness-aware: rules that can't be evaluated are skipped, weights are redistributed, and confidence drops. A prediction shouldn't look equally trustworthy when part of the evidence is missing.

Started as a side project with a deliberately small brief: take an LLM flow and see what went in, what happened at each step, what came out — and where it failed.

Most of the design effort went into keeping it small. I dropped the original "Auditor" concept when it started pulling the product toward evaluations and compliance, asked myself the uncomfortable "isn't this just Langfuse?" question early, and built incrementally: a Python tracing core first, kept separate from the API and inspector UI that grew around it later.

One detail tells you what kind of tool this wants to be: failed runs are persisted before the exception is re-raised. Every run becomes structured step-level traces — inputs, outputs, timing, status, errors — with recursive redaction on the write and read path. Losing the crash is the one thing a debugger isn't allowed to do.