Work
Enterprise platforms in production and a public portfolio of AI applications for regulated decisions — each one explicit about what is AI, what is code, what is checked and where a human decides — plus the programs I teach and the board I advise.
A calibrated credit model, notices written by code from a reviewed template, an independent eight-check verifier proven by mutation testing, and a human who makes the call. A read-only AI analyst explains cases — and code checks every number it states.
A calibrated risk model decides who is flagged, fairness is audited for every subgroup, an AI agent assembles the evidence, and a clinician makes every enrolment decision. It also reproduces how training on cost instead of illness burden skews who gets care.
An AI investigator for SEC Form 4 insider filings. It gathers evidence, drafts a finding, checks every claim against what it actually retrieved, and hands a clear, flag or escalate outcome to a human — while openly reporting where its accuracy still falls short.
An assistant that explains a patient's own record, with independent safety checks before and after the AI. Dosing questions are refused, emergencies get urgent guidance, and answers that conflict with the record — such as a critical lab value — are overridden and sent to a clinician.
A conversational prescription-refill assistant built on a 10-state workflow and a LangChain agent with 16+ tools. Controlled substances, expired prescriptions and unknown medications are routed to a doctor or pharmacist instead of being handled by the AI.
An agentic RAG system that validates an auto-insurance claim against policy records, retrieves the relevant policy language, checks the claimed repair cost against market estimates, and returns a structured, explainable coverage decision — through two interchangeable agent frameworks.
An investment-research prototype that pairs Yahoo Finance metrics with AI-strategy insights retrieved from company reports through RAG. Every answer is scored for groundedness and relevance by an LLM judge, and companies can be compared and ranked on both lenses.
A LangGraph RAG workflow that grades retrieved documents, rephrases the question when nothing relevant comes back, and checks each answer for hallucination before returning it — with caching and parallel retrieval to keep it fast.
An AI-assisted job-search tool that turns a résumé into targeted search queries, collects and enriches job postings, cleans and classifies them with LLM chains, and offers a career-focused chat assistant — with each model task routed to the provider that suits it.
A dual-model machine-learning system that turns raw trip telematics — GPS, accelerometer, speed — into 50+ driving-behavior features, then predicts both the probability and the likely severity of an insurance claim, with interpretable risk factors and risk tiers.
Agentic workflows that turn unstructured HVAC and refrigeration work orders into structured diagnostics, cost buckets and asset matches.
98% cost-bucket accuracy
A LangGraph agentic workflow that replaced manual expense coding on regulated PII data and discovered a governed taxonomy along the way.
99.76% classified at high confidence
A 22-repository GCP platform forecasting SKU-store sales and promotional redemption for ~350 stores over an 89-week horizon.
~50% lower WMAPE vs. legacy
Price-elasticity forecasting and a constrained optimizer that recommends multi-week discount schedules merchants adopt as-is.
+11% markdown revenue (latest benchmark)
Response-curve forecasting plus mixed-integer optimization that decides how much shelf each product section gets in every store.
−86% runtime (9 h → 75 min)
A 21-day, hourly forecast of orders and units for every fulfillment-center and store pair, feeding staffing and capacity planning.
~14% WMAPE in staging
Forecasts of outbound and pallet selection for 12 distribution centers, used to plan labor and transportation a month ahead.
9.05% WMAPE, flat across 28 days
Building single- and multi-agent systems in Python, from LLM and retrieval foundations to agents that plan, use tools, coordinate and can be evaluated and governed.
Skills I teach
AI agents for business teams: designing agents that plan, use tools and remember, applying them across finance, HR, retail, healthcare and customer service, and judging which use cases are worth automating.
Skills I teach
Production LLM applications end to end: prompting and its optimization, retrieval, fine-tuning, multi-agent systems, and knowing which approach fits a given problem.
Skills I teach
Serve on the advisory board of a recovery-science wellness startup, advising the founding team on technology strategy, platform architecture and AI/ML strategy across the product.
In the lab
Agents and tooling that extend the same patterns to legal, telecom, logistics and IT operations. Each has its own evaluation harness and CI; public demos are coming.
The shared library behind the portfolio's agent services: guardrails (injection screen, grounding, budgets, least-privilege projection), observability, per-node cost tracking, rate limiting and an eval CLI. Every abstraction had two real consumers before it became one.
View repositoryGrounded rental-law reasoning over Texas and California statutes plus a synthetic precedent corpus. It cites, verifies every citation, and refuses questions outside its scope.
View repositoryTelecom customer support that assumes the user might be hostile: input and output guardrails, an identity gate, redacted memory and specialist agents under a supervisor that checks grounding and scope.
View repositoryRoutes logistics exceptions — re-route, expedite, claim, cancel — with every recommendation traceable to a versioned rulebook, and a human review queue for high-stakes or low-confidence decisions.
View repositoryAn on-call assistant that retrieves logs, runs sandboxed diagnostics and drafts incident reports, with an evidence-and-honesty guardrail on every conclusion.
View repositoryA supervised coding-agent orchestrator: a thin supervisor decomposes a task, delegates to a coder model, and gates each change through checks and review before it is pushed.
View repositoryRohlik Orders Forecasting Challenge
Kaggle solution forecasting grocery delivery orders across Rohlik warehouses in Europe, for workforce and logistics planning.
MMA Fighter Analysis & Ranking
Scraping, rating and visualizing MMA fighter performance as a ranking tool for fans and analysts.
Open to consulting, advisory roles and select projects.
Tell me about the decision you're trying to improve. I'll tell you honestly whether AI is the right tool.