Applied AI engineer / New York
I build the systems that turn AI into
production work you can trust.
Four systems. Sole architect, engineer, and operator.
From the first API call to the failure mode at midnight.
Low-risk actions can run inside budget caps. Medium- and high-risk actions wait for approval. Mutations default to dry run.
Run programmatic style checks, then a second model pass that checks each product claim against retrieved evidence. Unsupported claims require a redraft.
Postgres row-level security enforces access. Automated tests attempt cross-tenant reads and writes across 18 tables.
A fail-closed hook checks commands against a policy the agent cannot modify. Unknown and forbidden actions are blocked.
Selected work / 01–04
Architecture you can inspect.
Outcomes you can question.
Google was optimizing for calls. Only 5% were qualified. I built the loop from transcript classification to keyword attribution to verified bidding outcomes.
Indexed outcomes · baseline = 100 · not a time series
Risk-tiered execution. Verified API mutations. Rollback for every mutation class.
Around-the-clock research, outreach, and inbound quoting. Product claims come from retrieved catalog data, with validation before send and human approval on quotes.
Product specifications
Pricing & lead time
Supporting documents
Separate drafting from permission to send. Unsupported facts trigger a redraft.
124K attribution rows connect marketing to client economics. Directors work their own leads. Finance gets the roll-up. Access is enforced in Postgres.
Test forbidden reads and writes across every tenant boundary. Protect attribution fields from edits.
A Python supervisor and Electron desktop app. Each task gets an isolated worktree, a durable event record, and a reviewable branch.
isolated worktreeisolated worktreeisolated worktreeThe agent cannot edit its own policy. Autonomy grows on evidence and is instantly revocable.
Also built / Website engineering
128 pages. 384 screenshot comparisons. 12,629 links checked.
Built and verified ahead of cutover.
The operating principles
Seven years building for marketing, sales, and finance. Three years on LLM APIs. The same engineering habits show up in every system.
How I workSeparate a proposed action from permission to execute it. Put approval where the consequence demands it.
Read back mutations. Test tenant isolation. Measure against a holdout. A green log is only a starting point.
Reviewable branches, incident records, runbooks, and controls that survive the person who wrote them.
Let’s build something that holds up.
I can walk you through the architecture, evaluation, and failure handling behind these systems.