Area
Applied AI & engineering leadership
On MyCareer.NJ.gov, I designed a shared AWS Bedrock foundation and a portfolio of applied-AI proofs of concept. The infrastructure runs in the state environment. The resident-facing features were not launched, and I do not claim resident outcomes from them.
Record
AI career guidance, accountable to evidence
I shipped the shared safety infrastructure behind MyCareer.NJ.gov's applied-AI portfolio, then built an unlaunched career-coach POC. Executed QA improved sharply and still exposed pilot-blocking gaps.
Open source
Fare Policy Assistant
Maturity: Deployed beta · reference implementation
A retrieval-grounded assistant that answers rider questions about fare and reduced-fare policies for a dozen-plus California transit agencies, built so its public evaluation harness has something to grade. Merge-blocking suites cover citations, refusals, forged conversation history, and English-Spanish parity, with versioned prompts and a committed regression baseline, and properties the harness has not measured are reported as not measured, never as passing. Ingesting real fare pages keeps surfacing findings: a current fares page still listing a pass product retired weeks earlier, and an inter-agency transfer published as good for 60 minutes on one agency's site with no window published on its partner's. It never determines anyone's eligibility.
plumbline
Maturity: v0.2.0
A fail-closed evaluation harness for government-facing chat systems: reproducible, provenance-stamped audit verdicts, byte-identical across Python versions, with no third-party dependencies. A silent or unreadable target scores zero rather than passing by absence, which is the defect an earlier version had: 174 empty responses once scored a perfect mark on every check phrased as the absence of a bad thing. The bundled dataset is synthetic and measures nothing about any real system; the harness is the product.
gauntlet
Maturity: v0.1.0 · on PyPI as gauntlet-evals
Merge-blocking evaluation gates for generative-AI features: YAML suites run against any HTTP endpoint or Python callable, fail the build on a miss, and emit a diffable JSON pack plus a reviewer document cross-referenced to California’s published GenAI risk framework. Aligned to that framework, never approved by it; the State has not reviewed or endorsed anything it emits. It evaluates a deployed feature in context, not a model, and it cannot verify a dishonest target.
cairn
Maturity: v0.3.0 · reference implementation
A grounded-or-silent reference assistant for public agencies: it answers only from a corpus the operator supplies, cites every claim, and refuses plainly when no source clears the threshold, with no model and no network at answer time. English, Spanish, and right-to-left languages. Its own front page leads with the configuration that once let an unsourced answer through, and the test that now fails without the fix. A demonstration of correct behavior to read, not a product to deploy.
Questions I get asked
What applied-AI systems did you design for MyCareer.NJ.gov?
I designed six primary workstreams: a grounded career coach, semantic search, résumé and skill matching, career intelligence, notifications, and content and accessibility automation. The broader portfolio included thirteen secondary proofs of concept built on a shared Bedrock foundation.
Which parts reached production?
The shared infrastructure runs in the state environment. The career coach and other resident-facing concepts remained proofs of concept awaiting a state pilot. No resident used the coach, so there is no resident outcome data and it should not be described as a live MyCareer.NJ.gov feature.
How did you test the career-coach boundary?
The proof of concept went through six pre-pilot reviews, about eighty adversarial cases that blocked merges, and a 210-answer known-answer evaluation. I tested prompt injection, unsupported programs and prices, missing data, topic drift, unsafe certainty, and English and Spanish response behavior. Those tests measure the prototype against a defined suite. They do not make a language model infallible.
What did the fifty-state scan establish?
My time-bounded scan of public workforce services in all fifty states did not surface an equivalent resident-facing conversational coach grounded in program records. That finding oriented the prototype. It is not proof that no comparable work existed privately or appeared later.
What does this site's assistant demonstrate?
It applies the same rule on a smaller public surface: retrieve reviewed material, answer only when the source supports the response, and stop when it does not. The source, server-side model call, refusal behavior, and production tests remain separate so each boundary can be inspected.