Portfolio
A few projects where I turned the reliability-first idea into working systems — from AI memory tools to financial-crime automation to new model architecture.
A local-first, markdown-native system that turns scattered workshop conversations into structured, retrievable knowledge. Its voice copilot, Spotter, captures structured notes live; an always-on "room ear" transcribes the session and answers questions from it; and a hand-built 3D knowledge graph, The Grind, surfaces IDEO/Lean synthesis you can export. No database, no lock-in — everything runs on your machine.
A multi-agent system that brings generative-AI speed to Suspicious Activity Report drafting without the audit risk. Built on my C.A.R.E.S. framework, it forces enumeration of every laundering typology, halts in code when a case needs human sign-off, and routes review across different AI model families to break shared blind spots. In a live side-by-side it produced audit-grade traceability — 83% vs. 41% legibility — where a naive pipeline silently bypassed the safety gate.
A book and a patent-pending framework (C.A.R.E.S.) for building AI agents that stay reliable after the demo is over — embedding the safety disciplines of nuclear power, aviation, and surgery directly into how agents operate, enforced in code outside the model so they can't be shortcut.
A novel transformer architecture I invented that builds an independent reliability check into the network itself — inspecting each internal update before trusting it, the way a reviewer checks work before it ships. Trained from scratch, it achieved the lowest error of every standard method it was tested against, on every single run.
A working discrete-event simulation of an emergency department — triage, fast-track, labs, beds, discharge — with a live dashboard for wait times, queue lengths, and staff and bed utilization, plus one-click stress tests. The kind of capacity-and-staffing analysis that normally needs specialist software and a simulation engineer, delivered as a single browser file.
A local, browser-based workflow tool that compresses the messy first phase of a consulting engagement — research, opportunity-mapping, and deck-drafting. It guides a consultant through a fixed public-research checklist, then synthesizes the findings into an executive-ready pitch-deck blueprint and an AI Opportunity Map.
A playbook and ROI model for operationalizing AI — a structured AI-operations approach paired with a calculator that turns assumptions about adoption, effort, and cost into a defensible return estimate, so leaders can decide where AI actually pays off before committing budget.
A local-first "people memory" app that keeps track of colleagues and contacts — rich profiles, photos, and dated notes — and even surfaces AI talking points before a conversation. Private by default and stored on your own machine: no server, no account. Built as a desktop app with Python and FastAPI.
An open-source alternative to NotebookLM's podcast feature. Point it at a webpage, PDF, YouTube video, or raw text and it writes a natural, two-host conversation about the material, then voices it with your choice of TTS provider — OpenAI, ElevenLabs, Google, or Edge — in short or long form and 100+ languages. Output is a finished MP3 or a transcript.