Product Leader.Founder.Builder.
I turn ambiguous problems into real software — designing AI workflows, marketplaces, and recruiting technology that compound over time.
Focus areas
What I've Built
The products, the problems they solve, and the decisions behind them.
Ryger
Founder · Case StudyA recruiter-owned talent network, discovered one workflow problem at a time.
Ryger didn't start as a recruiting product. It started as a question I kept running into during my own job search — and every answer changed what it should be. What follows is the sequence of discoveries that turned a matching tool into a network.
The discovery arc
- 01Initial Observation
- 02Research
- 03Workflow Discovery
- 04AI Trust Discovery
- 05Recruiting Nuance Discovery
- 06Marketplace Discovery
- 07Network Discovery
VeraTeach
LiveA production AI teaching workspace that carries a lesson from standards through Google Classroom to returned grades.
- Problem
- A teacher's week is spread across disconnected tools: one place to plan, another to look up standards, another to build the quiz, Google Classroom to hand it out, a spreadsheet to grade it, and a blank document whenever last year's material needs adapting. Each handoff is manual re-entry. Most AI education tools generate a lesson and stop there, dropping a plausible document into the middle of a workflow they don't participate in.
- Why I built it
- The workflow became the product. Lesson generation is one capability inside a system that spans the whole arc: a lesson is grounded in a 100,000-record standards corpus through retrieval rather than recollection, shaped by guidance the teacher writes before generation runs, extended into quizzes, worksheets, study guides and exit tickets, published to a class as a native Google Form where that fits, and graded in a workspace that scores the objective questions automatically and proposes — never applies — scores for written answers. Grades go back to Classroom. Teachers can also import material they already have rather than starting from a blank page. One orchestration layer coordinates every AI stage, and every AI write is previewed before it counts.
- Lesson
- The engineering that matters lives in the coordination: hybrid standards retrieval that filters, ranks by vector similarity, then reranks; immutable, trigger-written version history that makes every change reversible; copy-on-write scheduled instances so one lesson runs in many classes without collisions; cost- and stakes-based model routing; idempotent publishing into Google Classroom; and a grading path that keeps the teacher as the one who decides. AI earns its place by being orchestrated into real work and keeping the teacher accountable — not by making the call on its own.
Architecture
- Stack
- Next.js 16React 19TypeScriptSupabase / PostgresVercel
- Data & standards
- pgvectorMulti-state standardsAuthoritative ingestImmutable trigger-based versioning
- AI orchestration
- One orchestration layer across all AI stagesModel-agnostic routing via AI GatewayCost / stakes-based model routingRetrieval where correctness mattersPreview-first, teacher-approved writes
- Google Classroom
- Verified OAuth with encrypted tokensPersistent period → course mappingNative Google Forms quiz creationPublishing as Docs · Forms · live linksIdempotent re-publishGrades returned to Classroom
- Production
- Stripe subscriptionsUsage entitlementsAuthentication & account recoveryError monitoring
- Key Systems
- Hybrid standards retrieval (filter + vector + rerank)AI orchestration across stagesInstructional material generationAssessment → publish → grade → return loopAutomatic objective scoringAI-assisted free-response suggestionsLesson import from existing materialCopy-on-write scheduled instancesDestination-abstracted publishing
Resume
Resumes grounded in your whole career, not just one role
- Problem
- Resumes are built for a single role, so experience that doesn't match the current application quietly gets cut. Most tools make it worse — optimizing for keywords, or inventing accomplishments that no longer sound like you.
- Why I built it
- Built to tell a more complete career story. Resume pairs a traditional resume with a deeper career narrative, then reads a job description against both — surfacing the most relevant real experience first, always grounded in your actual work history.
- Lesson
- Trust is the product. Every resume has to stay truthful, free of invented metrics, and still sound like the person behind it — even though a model does the drafting.
Architecture
- Frontend
- Next.jsTypeScriptTailwindVercel
- Platform
- AWS LambdaAPI GatewayDynamoDBS3SES
- AI
- OpenAI
- Identity
- Custom authentication
- Key Systems
- Career story engineResume generation workflowsGrounded content generationJob alignment workflowsTruth-preservation controls
ReadMyStrip
AI-Powered Water Test Strip Reader for iPhone
- Problem
- Testing pool, spa, or aquarium water means reading a strip of subtle color pads against a reference chart — slow, subjective, and easy to get wrong. Results shift with lighting, eyesight, and guesswork, right when accuracy matters most.
- Why I built it
- ReadMyStrip turns that into a single photo. I led the product from concept through production release — product strategy, UX, AI workflow design, engineering, backend architecture, and App Store launch — so anyone can point their iPhone at a used test strip and get reliable, interpreted results in seconds.
- Lesson
- Reliability is the product. A technically complex image-analysis workflow has to land as one trustworthy tap — onboarding, edge cases, subscriptions, and Apple's review process all handled so the user never has to think about them.
Architecture
- App
- ExpoReact NativeTypeScriptiOS
- AI
- Image analysis workflowTest strip interpretation
- Platform
- AuthenticationBackend APIsCloud storage
- Payments
- RevenueCatSubscription management
- Key Systems
- Single-photo capture flowAI water-chemistry interpretationEnd-to-end onboarding to resultsSubscription managementApp Store production release
Where I've actually done this.
Recurring patterns across the products I've designed and built — not a skills list. Each competency names the work, then points to where it shows up, so it's clear these are habits, not one-off implementations. Every one is something I can walk through in detail, tradeoffs and all.
Retrieval-Augmented Generation (RAG)
Ground generation in authoritative sources instead of model memory, so outputs are correct and defensible.
- VeraTeachHybrid retrieval over authoritative academic standards — a hard metadata filter, then vector similarity, then an AI reranker — feeds the model a bounded, correct candidate set.
- RygerEvidence pulled from a candidate's real history and attached to each job requirement, so a match is reasoned over retrieved proof rather than recalled.
Vector Search
Search by meaning, not keywords — with guardrails so relevance never overrides correctness.
- VeraTeachpgvector embeddings rank standards by semantic similarity, but only within a hard-filtered set, so meaning-based search can't return an out-of-jurisdiction result.
- RygerMoved matching past keyword presence — a resume listing “Agile” never proved the experience — toward relevance judged on demonstrated work.
Model Evaluation
Measure output quality with rubrics and repeatable checks instead of eyeballing a few good demos.
- VeraTeachRubric-based, evidence-grounded evaluation scores lessons against explicit criteria; a regression harness of scripted checks guards quality as prompts and models change.
- RygerMatch quality judged at the requirement level with supporting evidence, not a single opaque score.
Model Comparison & Routing
Pick the right model per task by weighing quality, latency, and cost — not defaulting to the biggest one.
- VeraTeachPolicy-driven, model-agnostic routing across Claude Sonnet, Claude Haiku, and GPT tiers — cheap models for low-stakes materials, stronger models reserved for summative tests.
- ReadMyStripChose a vision-capable model for on-device-quality strip reading, balanced against latency and per-call cost for a consumer app.
Structured Outputs
Constrain generation to typed schemas so downstream systems get reliable data, not free text.
- VeraTeachEvery production AI call returns schema-validated output; the model can't emit free-form text or pick a value outside the allowed set (it can't invent a standard).
- ReadMyStripWater-chemistry readings returned as structured, typed results the app can render and act on directly.
Prompt Orchestration
Coordinate many AI steps into one reliable workflow rather than one brittle mega-prompt.
- VeraTeachA single orchestration layer sequences retrieval, generation, review, adaptation, evaluation, material creation, and publishing — each a discrete, testable step.
- RygerMulti-step match-reasoning workflows chained behind a single recruiter action.
Grounding
Tie every AI claim to real evidence to reduce hallucination and keep outputs trustworthy.
- VeraTeachStandards suggestions are constrained to retrieved candidates — the model cannot fabricate a standard that doesn't exist.
- ResumeTruth-preservation controls keep resumes free of invented metrics and grounded in real work history.
- RygerEvery match recommendation carries the evidence that supports it, per requirement.
Human-in-the-loop
Keep AI assistive — the human reviews and approves before anything counts.
- VeraTeachPreview-first: AI engines persist nothing; teachers review, edit, and approve before any write, with full version history behind them.
- RygerAI clears the obvious matches and misses; the recruiter keeps the nuanced calls where judgment adds value.
AI Reliability
Make AI dependable in production — retries, timeouts, validation, and reversibility.
- VeraTeachA centralized execution seam wraps every call with retry, timeout, and schema validation; immutable, trigger-written version history makes every AI change reversible.
- ReadMyStripA technically complex image-analysis workflow made dependable enough for one-tap consumer use and App Store review.
Cost Optimization
Control the unit economics of AI features without sacrificing quality where it matters.
- VeraTeachLow-stakes materials routed to a model at roughly a third of the cost; premium models spent only where correctness is critical. Cached standards suggestions avoid repeat embedding and rerank cost.
Workflow Design
Start from how the work actually flows, then fit software and AI into it.
- VeraTeachThe product follows a teacher's real arc — planning through classroom delivery — instead of dropping a generated document into a vacuum.
- RygerRedesigned candidate review to end the constant context-switching between a job description and a resume.
- ResumeReads a job description against a full career narrative to surface the most relevant real experience first.
Product Discovery
Find the real problem through research and use before building — and let it reshape the product.
- RygerA seven-discovery sequence turned a matching tool into a recruiter-owned talent network.
- VeraTeachDiscovered teachers don't need lessons generated so much as lessons that adapt — reshaping the product around adaptation.
- ReadMyStripReduced a fiddly, error-prone testing routine to a single photo.
Principles over playbooks.
A consistent way of thinking about product problems — the lens I bring before any specific tool or feature enters the conversation.
Start with the workflow, not the technology.
Understand how the work actually gets done before deciding what to build. The best technology disappears into a better workflow.
Optimize for outcomes, not activity.
Shipping features is easy. Moving the metric that matters is the job. I measure progress by the problem getting smaller.
Reduce friction before adding features.
Most products are slowed by friction, not missing capability. Removing steps usually beats adding them.
Build systems that compound over time.
Favor decisions that get more valuable as they accumulate — data, relationships, and structure that pay off later.
Make AI useful, explainable, and practical.
AI should earn trust by being transparent and reliable in real work — not impressive in a demo and fragile in production.
Why I approach problems the way I do.
Over the last 15+ years I've worked across customer service, team leadership, process improvement, process management, enterprise risk, data governance, and product management — inside banking, risk, operations, and product organizations.
The same problems kept recurring: workflow design, operational efficiency, stakeholder alignment, governance, and reducing friction in complex processes. Many of the themes in my projects — trust, reducing context switching, preserving valuable work, and building systems that scale — originated long before I started building software.
Career highlights
Capital One
Product Management & Enterprise Risk
Built and managed products supporting governance, compliance, risk management, enterprise data workflows, and operational decision-making — focused on product strategy, roadmap development, stakeholder alignment, and large-scale enterprise initiatives.
PNC Bank
Process Improvement Consulting
Partnered with business leaders and operational teams to identify inefficiencies, improve workflows, coach teams, and drive continuous-improvement initiatives.
TD Bank
Process Management & Operational Leadership
Progressed from frontline leadership into business process management and enterprise process improvement — leading large-scale process optimization, data analysis, operational redesign, and organizational improvement programs.
The projects demonstrate what I build. My experience explains why I think about problems the way I do.
Building something interesting?
I'm always glad to trade notes on products, marketplaces, and making AI genuinely useful. The best way to reach me is below.
Open to Senior AI Product · Technical Product · AI Platform PM roles