Projects

Projects & Case Studies

Every project here ran against live business data rather than a tidy sample set. Client work is anonymized; the measurements are not.

Client work September 2026

Insights from a COO's Inbox

A client asked where AI could help their operations. I turned twelve months of their COO's mailbox into a counted table of recurring problems, each carrying the threads that evidence it. Two of the thirty-two called for AI.

1,290
threads read and labeled
32
recurring problems
2
of 32 called for AI
Python DuckDB Gmail API LLM extraction Taxonomy Read the case study
Client work July 2026

Testing Whether an LLM Improved Bid Estimates

A contractor wanted an LLM to read field notes and sharpen their labor estimates. I built it, measured it against plain statistics, and the statistics won. The systemic check shipped, the per-line risk score did not.

96%
of jobs one code overran
64%
of excess hours caught
1.1x
lift, so it didn't ship
Python SQL LLM evaluation Statistics Backtesting Read the case study
Client work September 2026

Catching Product Data Drift Across Four Sales Systems

Four systems each kept their own record of the same product, and nothing enforced the link between them. When it broke, no error fired anywhere and a person found it 34 days later. A nightly check now reports it the next morning.

34 days
to notice, before
1 night
to notice, after
6
finding categories
Python Postgres Shopify API Amazon SP-API Data quality Read the case study
Platform Since February 2026

How I Sandbox LLM Agents That Read Production Databases

Scheduled agents read production databases every night and report what they find. Each run gets a throwaway network holding only the one database its policy names, and a database login with no ability to write.

47
containers
1,049
archived runs
0
write grants
Docker Claude Postgres Prometheus Security Read the case study