Selected Work · Yellow Labs Products · January 2025 – Present

Production products across recruiting, education and consumer apps.

Each product exposed a different problem in designing AI systems people can rely on: grounding generation in authoritative sources, evaluating quality before it reaches users, deciding what a model should never decide, and turning messy real-world input into answers people act on.

I use AI-assisted development to work directly in the implementation layer, while owning the product decisions, architecture, evaluation criteria and production validation.

VeraTeach

Live · Web · K–12

An AI teaching workspace that carries a lesson from state standards through Google Classroom to returned grades.

Multi-stage AI orchestration
Shared 110K+ standards layer
Evaluation and model routing
Human-in-the-loop grading

Key decision: AI drafts grading suggestions; only teacher-approved scores go back to Classroom.
Signal: Rubric evaluation went from 74s to 19s at ~2.1× lower cost, with near-identical scores.

Read case study verateach.com ↗

Grocery Sidekick

Live · iOS

Scan a grocery product to see which of the allergens and ingredients you watch for it contains, and which ingredient is responsible.

Large-scale data normalization
Ingredient resolution
Deterministic classification
Benchmarks and audits

Key decision: Turn messy source data into a reliable product layer: inconsistent label records normalized into stable product identities, ingredients resolved against a curated dictionary, and classifications computed deterministically at build time.
Signal: 400,000+ foods and 430,000+ barcodes bundled on the phone, so lookups work with no signal.

Read case study App Store ↗

Ryger

Live · Web

Recruiting intelligence with evidence as the boundary: it structures candidate experience, matches it against role requirements, and generates targeted resumes limited to what the candidate can actually support.

Requirement-level matching
Evidence-grounded resume generation
Shared candidate-intelligence layer
Discovery that reshaped the thesis

Key decision: Evidence is the boundary. Each requirement is Met, Review or Missed with its supporting evidence, and a generated resume line whose numbers, dates or employers aren’t in the source is removed rather than trusted.

Read case study ryger.app ↗

ReadMyStrip

Live · iOS · Paying customers

Photograph a pool or spa test strip next to your own bottle’s chart and get every reading in seconds.

Vision-model pipeline for real-world photos
Deterministic validation of model output
Subscription monetization
Iteration from production feedback

Key decision: The model only perceives; code decides. A reading that isn’t printed on your chart becomes “unknown,” never a guess.
Signal: Paying customers — and production feedback led to scan history and notes.

Read case study App Store ↗

Capabilities by product

Where each capability shows up.

The four capability themes from the homepage, mapped to the work that evidences them — including the enterprise work at Capital One.

Theme	Capital One	VeraTeach	Grocery Sidekick	Ryger	ReadMyStrip
AI product systems	—	Retrieval-grounded generation; teacher approval	—	Evidence-grounded matching and resumes	Model perceives, code decides
Evaluation & AI quality	—	Blind quality / latency / cost comparison; rubric	—	—	Image-quality gates
Data & decision intelligence	Data-management process ownership	185-framework standards corpus	400K+ foods in a deterministic data layer	Structured evidence; Met / Review / Missed	—
Product leadership & delivery	Cross-functional delivery; 57 → 2	The workflow became the product	—	Discovery → talent network	Features from customer feedback