Six projects from the last twenty-four months.
Selected work, with the receipts attached. Each one was a fixed-scope engagement ending in a system in production, an on-call rotation that doesn't keep anyone up at night, and a team that could maintain it after we left. The names we'll share under NDA.
SHRINKAGE RECOVERED / YR
An anomaly detector that found the leak nobody was looking for.
A regional 3PL was losing inventory in a pattern their existing BI tools never surfaced. We built a streaming feature pipeline, layered an isolation-forest ensemble on top, and tied the alerts to a paging system that routed by route, lane, and shift. The system flagged recurring shrinkage at three of seventeen hubs in week two. Quietly paid for itself in eleven.
FEWER ESCALATIONS
A clinical RAG assistant on 14,000 pages of internal SOPs.
Hybrid retrieval, a cross-encoder reranker, and a strict citation contract. Every answer linked to the section it came from. Wrong answers measurably down. Time-to-answer down 6×. Nurses stopped paging the on-call clinician at 2am to look up a workflow.
UNPLANNED DOWNTIME
Predictive maintenance on a fleet of 1,200 industrial assets.
Sensor ingest from 11 site PLCs, a Postgres-backed feature store, gradient-boosted models per asset family, and a quiet maintenance dashboard. The team running it can retrain weekly without our help, which was always the point.
FASTER ONBOARDING
An onboarding platform that retired four spreadsheets and a Zapier graveyard.
A multi-tenant admin platform with workflow orchestration, role-scoped audit logs, and a customer-facing portal. New customer onboarding went from six days to four hours. The customer success team got their Mondays back.
FIELD-LEVEL ACCURACY
A document extraction pipeline for 38 lender statement formats.
OCR plus layout-aware extraction plus an LLM correction pass, with a human-in-the-loop queue for the long tail. Replaced eleven hours of manual data entry every day. The audit trail kept compliance happy on the first review.
QUERYABLE IN SECONDS
A geospatial data layer that answered 340TB of questions in seconds.
Parquet on object storage, DuckDB for analyst queries, a thin tile server for the front end. Their researchers stopped exporting CSVs at 9pm to see results at 7am the next morning.
Where the signal usually comes from.
The model paged us before the customer did. That single page paid for the whole engagement, and we still tell that story in board meetings.
They scoped a 10-week project. They delivered in 9. Then they stayed for 2 more on us to write the on-call docs nobody else would have written.
The thing I appreciate most. They argued with us when we were wrong. The thing I respect. They could prove it.
Want the full write-up? Numbers, code, names.
We can share a full case study under NDA: stack, architecture, eval results, the post-incident learnings, even the parts that didn't work. Tell us which one you want and which problem you're trying to solve.