Timeline

Growth over titles.

What I built, learned, got wrong, and changed — each phase includes the mistake, because that's where the learning is.

Read the résumé

Independent AI products
2025 – present
Built
Four live AI products on a $0 stack — a NL→SQL agent (DBWhisper), an honest backtester (TradePulse), a +EV sports-ML system (CrownWager), and a multi-user AI chat platform (LLM Studio).
Learned
Safety-first LLM systems: deterministic gates around model output, retrieval over context-stuffing, and evaluation before pixels.
Mistake
Early on I trusted model output where a deterministic gate was needed — an LLM writing SQL against a live database with only a prompt between it and a mistake.
Changed
Fail-closed validators and 'look-ahead-free by construction' became defaults, not afterthoughts — the safety is now a property of the architecture.
AI/ML Engineer
Sevina Technologies
2024 – present
Built
Production healthcare-AI automation — clinical-compliance and reimbursement pipelines (MDS/PDPM), LLM document analysis with multi-provider routing and RAG, and eligibility-verification services, Dockerized and CI'd.
Learned
Building AI under real regulatory constraints (HIPAA): synthetic data, aggregate metrics, auditability, and knowing when the boring deterministic approach is the correct one.
Mistake
I leaned on an LLM for correctness that belonged in deterministic code — enumerable rules an auditor needs to reproduce.
Changed
Rules and validation moved into code the model can't override; the LLM assists, it doesn't adjudicate.
Junior Python Developer
Linescripts Software
2022 – 2024
Built
End-to-end web applications and RESTful services — a Hospital Management System, e-commerce platforms, and third-party API integrations, across Python, Django/DRF, and a JavaScript frontend.
Learned
How to ship maintainable software end to end, and that the interesting part is usually everything around the feature — the tests, the errors, the edges.
Mistake
I optimized for shipping features over tests, and paid for it later in regressions I could have caught.
Changed
CI and tests became non-negotiable before a feature is 'done' — the definition of done now includes proof it works.