Technical Projects

Selected data and software projects. Analysis tools, detection platforms, and utilities built end to end.

Selected Technical Projects

Medicaid Claims Explorer screenshot
Medicaid Claims Explorer

Python, DuckDB, Streamlit, Plotly

The same dataset is charted on the Data Analysis page.

  • Interactive analysis of a 238-million-row, roughly 11 GB Medicaid claims dataset, comparing billing and servicing provider spending patterns, with peer benchmarking and provider-level summaries.
  • Columnar Parquet storage with input-validated, escaped SQL filtering over DuckDB.
  • Paired with a Jupyter notebook recording the full exploration and cleaning methodology.
Personal Portfolio Site screenshot
Personal Portfolio Site

HTML, CSS, JavaScript, Python, DuckDB, Git, Cloudflare Pages

  • Multi-page site (Home, Resume, Technical Projects, Data Analysis, Freelance, Creative) with a shared design system built from scratch in plain HTML, CSS, and JavaScript, no framework.
  • Interactive Data Analysis page backed by a Python and DuckDB pipeline whose output is automatically checked against HHS's own published figures, matching to the cent.
  • Directed every structural and content decision, page architecture, copy, ordering, and design choices, built with Claude Code through an iterative process.
  • Deployed through a Git-based pipeline to Cloudflare Pages with automatic rebuilds on push, behind a pre-commit check that blocks phone numbers, secrets, local file paths, and image metadata from being published.
Fraud, Waste, and Abuse Detection Demo screenshot
Fraud, Waste, and Abuse Detection Demo

Python, DuckDB, pytest

  • Working demonstration of a fraud, waste, and abuse detection platform that ingests any CSV or Excel file and profiles every column to infer semantic type.
  • Runs nine statistical detectors and ranks findings by severity with plain-language explanations.
  • Runs fully locally, so sensitive data never leaves the machine.
Ecology Simulation screenshot
Ecology Simulation

JavaScript, HTML Canvas, no dependencies

Run it live · Source (MIT) · runs in your browser, best on a desktop

  • I wanted to see whether a world I could imagine would do anything interesting if I built it. Small square organisms sense, eat, sleep, fight, flee and breed in a 2D world with a day and night cycle. Seeded, so any run can be replayed.
  • The design is mine: eleven inherited genes, senses that shrink at night, a behaviour priority stack, and homes that form at a well-fed sleep, drift toward where an organism eats, and get abandoned after attacks. There is no aggression gene; violence comes from hunger, relative power and opportunity.
  • Claude Code wrote most of the implementation to my direction. I designed it, tuned it and tested it until the ecology held together at any world size, about 900 organisms at 2 ms a tick.
  • Part art piece, part experiment in what AI can build when someone directs it with a clear idea. I'm still playing with it and watching what comes of it.
Software keyboard instrument with a string-section preset loaded
Software Keyboard Instrument

C++20, JUCE, CMake · Windows desktop app

A prototype I use myself, started October 2026. Six of seven planned milestones are built.

  • Plays six synth engines written from scratch, a sampled orchestra, and modeled pianos and organs live from a MIDI keyboard, with 115 presets.
  • Records performances as multitrack takes, edited in a piano roll (quantize, humanize, tempo map) and exported as MIDI, WAV, MP3, FLAC, or per-track stems.
  • Turns a folder of audio recordings into a playable instrument through pitch detection, trimming, loop finding, and key mapping.
  • I wrote the spec and milestone plan, made the product and design decisions, and play-tested every milestone. Claude Code wrote the code under my direction. 3,688 automated checks pass across about 27,500 lines of C++.
Driving game: a car on a desert town main street at night
Open-World Driving Game Prototype

Unreal Engine 5, C++, Blender with Python · playable in the editor, not yet released

A prototype I started in October 2026: drive an electric car through a desert town at night, with traffic and robots in the streets.

  • The town is generated: Python layout scripts place buildings, streets, and props, Blender builds the meshes, and Unreal assembles them with lit interiors, neon signs, and lot lighting.
  • A C++ traffic system drives 69 AI cars through 4 signalized junctions with stop signs, left-turn yielding, and honking. A C++ crowd system moves 206 robots and drones that cross on walk signals and dodge the player.
  • Engine, robot, drone, and ambient sound is synthesized live from vehicle speed, engine RPM, and player position.
  • I set the concept and creative direction, made the design decisions, and tested it by driving. Claude Code produced most of the code, 3D models, and level edits under my direction, working in the Unreal editor and scripting Blender.
GPU video editor with a demo timeline, effect stack, and preview
GPU Video Editor

TypeScript, Electron, WebGL2 and GLSL, Web Audio, FFmpeg · Windows desktop app

A desktop video editor I use on my own projects, started October 2026. All editing and exporting runs on the graphics card, fully offline.

  • A multi-track timeline with ripple editing, trimming, keyframes, speed changes, reverse, and freeze frames, previewed in real time on the GPU.
  • 19 picture effects written as custom shaders (color, glow, lens, CRT, film grain, LUTs), 12 transitions, titles, and a sound mixer with EQ, compression, reverb, a master limiter, and loudness normalization.
  • Exports to H.264, H.265, WebM, GIF, WAV, or MP3 through the same renderer as the preview, and 136 automated checks confirm the export matches the preview frame for frame. Export on a 20-second 1080p test project went from 3.2 to 43 frames per second.
  • I set the requirements and design and tested every step on my own footage before the next began. Claude Code wrote the code and its 328 unit tests under my direction.
Local AI news feed with rated stories and a system health panel
Local AI News Feed

Python, FastAPI, HTMX, SQLite, Ollama with Qwen models · runs locally on Windows

A personal news feed I use daily, started October 2026. It collects AI and tech news and uses a local AI model to sort and rate stories, with no cloud AI services.

  • Checks 47 public news sources on a schedule and merges coverage of the same event into one story, using an embedding model to find matches and the local AI to confirm borderline ones.
  • Places each story in a section, labels it news or discussion, and rates it 1 to 5 stars with a one-sentence reason, learning from my up and down votes. Major stories trigger a desktop alert.
  • I chose the models by testing them: on 20 hand-labeled headlines, the 9B model sorted 17 correctly versus 9 for a 4B model, and duplicate detection matched my judgment on 17 of 18 pairs.
  • I defined the requirements and design and tested it through daily use. It was built with AI-assisted development using Claude Code, and 138 automated tests pass.
Video Clipping Tool screenshot
Video Clipping Tool

JavaScript, Electron, FFmpeg

  • Desktop application for frame-accurate video trimming with waveform-based timeline navigation.
  • Automatic proxy-transcoding fallback for codecs the runtime cannot decode natively.
  • Packaged with an electron-builder Windows configuration.
Offline Text-to-Speech Studio screenshot
Offline Text-to-Speech Studio

Python, Gradio, PyTorch

  • Web interface unifying two speech-synthesis engines behind one local interface with a shared engine calling convention.
  • Kokoro-82M preset voices across nine languages and Chatterbox zero-shot voice cloning across 23.
  • Runs fully offline after the first-run model download.
Additional Tooling
  • Automated multi-camera video edit generation driven by user-supplied beat patterns and tempo maps, with automatic resolution and frame-rate conforming.
  • Batch media processing pipelines (FFmpeg, pydub).
  • Image compositing utilities and a browser-based color palette studio.

Curious what you could build with AI? I help people get started. Get help