lab
Small things and experiments
Work in progress, tools, and things I tried that did not go as planned. Failures are published with a write-up because that is where the learning is.
now — 2026-09-08
- Writing part 4 of the agents series — the trace chapter, which keeps growing.
- Rebuilding the fraud case-preparation agent on IEEE-CIS so the anchoring experiment is reproducible.
- Reading about constrained optimisation for price ladders. The pricing rebuild needs a better solver than the one I shipped.
Uplift sandbox
betaUpload a treatment/control CSV, get uplift deciles and the qini curve. Runs in the browser.
Updated 2026-08-30
Runnable
Eval harness for tool-calling agents
productionThe harness from the fraud work, generalised: trace capture, assertion library, regression report per commit.
Updated 2026-08-12
Forecast-to-policy calculator
prototypeType a MAPE and a lead time, see the working capital it implies. Deliberately crude.
Updated 2026-07-25
Runnable
Semantic layer from dbt manifest
prototypeGenerates a queryable metric layer from a dbt manifest. Works on three of the five projects I tried.
Updated 2026-06-18
Retrieval over meeting transcripts
failedAbandoned. Retrieval quality was fine; the problem was that nobody trusted a summary of a meeting they attended. No amount of model work fixes a trust problem. Write-up covers what I would test first next time.
Updated 2026-05-02