Work
My work.
I build and operate AI and workflow products inside estimating, quoting, booking, and private knowledge work, where a wrong output costs money or trust.
Each case below records who used it, the product bet, the quality bar it had to clear, what shipped, what changed, and the measured result.
-
01Read the case
AI reconciliation
The Review Desk
Reconciliation · evidence provenance · human-in-the-loop AI
A commercial bid was assembled from five documents that often disagreed with one another. Those inconsistencies were easy to miss, and expensive when they reached a customer.
I analyzed a decade of estimating lessons, turned the recurring patterns into a detection framework, and built a pipeline that reads the complete bid package, reconciles its inputs, and flags places where the bid appears to be repeating a known failure pattern.
Every flag includes the underlying source evidence. The estimator accepts or dismisses the finding before the quote is sent.
$221,915.70 surfaced · $3.44M reviewed · 6.4% verified value catch rate · zero validated findings escaped into a sent quote
-
02Read the case
Workflow automation
Quote Import
Deterministic workflow automation · state transformation · production desktop software
A roughly 50-person controls contractor quoted every job by retyping the same facts through three disconnected systems, twice at the seams. I took a deterministic browse-review-save tool from workflow discovery through beta, production rollout, and post-launch expansion across product lines, export paths, and pricing workflows.
30 to 60 hrs/wk returned to a 12-person estimating team · 12 releases · no rollbacks
-
03Read the case
Privacy-aware AI
Stem
Agent runtime · permission gates · evaluation infrastructure · local-first AI
A daily AI practice that mixed sensitive documents with ordinary questions defaulted every request to the same cloud model, or slowed down to sort them by hand. I built Stem around a fail-to-local router and a deterministic redaction gate, the one code path a request has to clear before it can reach a cloud model, and I operate the product end to end.
19,124 production events · zero confirmed leaks · 96/96 self-test gate
-
04Read the case
Agent orchestration
Stem Orchestrator
Dependency-aware parallel execution · controlled autonomy · measured throughput
Every task ran through one model in one growing context, one step at a time, whether or not the steps depended on each other. The orchestrator I built decomposes each task into a graph, fans out the independent pieces, and grades the merged result before it counts as finished.
2.25× realized end-to-end throughput on real decomposable work
-
05Read the case
AI evaluation
Pricing QA Gate
Deterministic validation · regression harness · human escalation
A fuzzy-matching path guessed catalog prices from ambiguous line descriptions and looked as trustworthy as the exact-match path beside it. I built and scored the fuzzy path separately, and the gate in front of it blocked the path before a single guessed price reached a live quote.
63 of 66 fuzzy-matched prices wrong · the path never touched a live quote
-
06Read the case
Structured automation
Quote Assembly
Deterministic transformation · structured outputs · first-pass validation
An estimator used to retype a raw vendor export into the company's quote format by hand, under deadline, on every job. The automation I built assembles the full quote from that same export in one pass and matched every priced line on its first validation.
24-line export to an 88-line import-ready quote · 20 of 20 priced lines matched on the pilot job · 91 of 98 on an unseen job
-
07Read the case
Transactional product
Studio Booking Platform
Transactional state · payments and webhooks · invariants under parallel load
A boutique pilates studio's hosted booking SaaS kept losing count of members' class credits, and the owner had no ledger to reconcile against. I built a custom booking and payments platform spine-first on the studio's own Stripe account and proved every capacity, credit, and webhook invariant under genuinely parallel load before designing a single screen.
21 of 21 concurrency invariants held on two back-to-back runs · stopped at a validated working prototype; the studio stayed on its existing SaaS
Also built: thirteen working products across iOS and the web, audio instruments, utilities, and games made with my kids. See the fleet.