THE WORK
Systems I've personally designed, built, and operated — AI delivery pipelines, 0→1 products, and the agentic workflows that run them. Not slideware. Shipped things.
FLAGSHIP // AI PRODUCT
SHIPPED · iOS + ANDROIDONLY EVER
An AI-native test-prep platform on iOS and Android. I own it end to end — the AI, the product, and the content engine — and lead the distributed team of engineers and designers who ship it.
- An AI that turns an official exam syllabus into study material and graded assessments — conversational, Bloom's-tiered, and scored against a rubric with human-in-the-loop QC.
- Source-grounded: every concept carries evidence and a confidence check, and every answer traces back to its source — the anti-hallucination discipline regulated prep demands.
- 36 exam curriculums across 7 categories, incl. FINRA SIE / Series 7 / 63.
- End to end: the generation pipeline (multi-provider, batch + vector DB), the Next.js product (~1,100 programmatic-SEO exam pages, spaced-repetition study workspace), and the content/GTM engine — plus a hands-on app redesign with Claude Code.
FLAGSHIP // AI SYSTEM
SHIPPED · 0→1AQUAINTEL
An AI GTM-intelligence system for companies that sell to U.S. water utilities. It discovers ~12,000 utilities, reads their public documents with AI, and surfaces which ones are about to spend.
- A cost-aware model cascade reads utility documents — RFPs, capital plans, budgets — and ranks real buying signals; cheap models filter, expensive ones escalate only on low confidence. A fraction of a cent per document.
- Scored honestly against a human-curated gold set, with a versioned rubric taxonomy and per-run cost metering — evals, not vibes.
- A Next.js product on top: a national utility map, account deep-dives, and a grounded “Ask AquaIntel” RAG chat over a local quantized vector store.
- A discovery pipeline enumerates ~12,000 utilities from federal EPA data — deterministic core, LLM only for the genuinely fuzzy steps.
MORE_BUILDS
RESOURCE PROJECTS
0→1 delivery at NRGI: led a team of developers and designers to ship resourceprojects.org — turned ambiguous stakeholder needs into structured requirements and shipped the alpha in three months. An automated R/SQL pipeline cut processing cost 20x and converted 16,000+ rows of non-machine-readable PDFs into structured data.
KINETICA
Partnered with product and engineering on positioning and launches for a real-time analytic database. Built the developer-education program from scratch — 10 courses, 50+ videos, 25 GitHub repos — and delivered a full rebrand at 1/10th the prior budget.
THIS_SITE
The meta-proof: this site runs an agentic writing pipeline — brainstorm → draft → blind voice review → publish — built from subagents, a versioned voice contract, and skills. Spec-driven, human-in-the-loop, eval-gated. Same patterns the roles ask for. (Click for confetti.)