Evidence
What you can check before you trust me with a client.
Client work is under NDA or waiting on permission to publish. So all of
this is public data I can show in full: method, code, and the results
that went against me.
I find out that something does not work on my time and my budget, before
it reaches your client.
8Live systems, publicly evaluated
21Measured results published
8Went against the build
~25Team I built and ran
Evaluated retrieval, NTSB public data
Black Box AI
Question routing, a guarded SQL path, retrieval evaluation and an audit trail across 7,462 public aviation safety reports.
Decision No clear win
Limit
Hybrid retrieval did not clearly separate from simpler methods, so the product defaults to semantic search and exposes all three. The pilot set of 17 hand-labeled questions is too small to support a method choice.
Read the case study → Decision optimization, synthetic benchmark
Dispatch Optimizer
Technician assignments under skills, capacity, deadlines, travel and overtime constraints, with both baselines and the optimizer built for comparison.
Greedy beat naive on 99% of runs
Limit
Synthetic data, not a production operation. The timed solver is nondeterministic, so the distribution matters more than any single scenario.
Read the case study → Statistical inference, public mortgage data
Fair Lending Lab
Five preregistered hypotheses on 41,287 curated applications drawn from 210,643 raw HMDA rows, with effect sizes, resampling and multiplicity controls.
Survived adjustment 44.4%
Limit
HMDA lacks key underwriting variables. These findings are screening signals for deeper review, not discrimination findings.
Read the case study →
Five more evaluated systems, including three where the result went against
the build.
View all eight and the evidence ledger →
Before any of it I ran the operations I was automating: a behavioral
health practice of roughly twenty five people, covering staffing, payroll,
billing, compliance, and the tracking systems that ran it day to day.
Earlier I replaced paper and spreadsheet case tracking at a social
services agency with database applications that real caseworkers actually
adopted. That is the part that does not show up in a code sample, and it
is why I ask what a report is for before I ask what it should look like.
There are no client logos on this page, and there is a reason. A
subcontractor who will show you someone else's confidential work is
telling you exactly what they will do with yours.