These are the projects I would pull up if we were sitting together and you asked what I have been working on. Some shipped, some taught me something, and some did both.
Not necessarily the biggest repos. They are the clearest examples of how I think: make the hard thing usable, keep the person close to the controls, and learn by building the whole loop.
Go, local LLMs, CLI automation
A shell interceptor and local-AI command agent that catches mistyped commands, suggests corrected terminal actions, and can operate in agent mode from the CLI.
Why I care about it
This is the kind of tool I wanted myself: mess up a command, get a better one, and stay in the terminal.
Python, Faster Whisper, AppleScript, vector search
A local macOS assistant that combines voice input, document retrieval, graph-style context, and AppleScript control.
Why I care about it
I wanted an assistant that could use my actual computer context without sending everything somewhere else.
Go, macOS, launchd, Docker, SSH
A native macOS operations HUD for scripts, launchd agents, Docker containers, SSH services, and remote jobs.
Why I care about it
I got tired of checking five places to see what was running, so I put the useful state and controls in one place.
Python, static analysis, local-first tooling
A local-first repository scanner that converts codebases into structured summaries while keeping source private.
Why I care about it
I can get oriented in an unfamiliar repo without uploading somebody else’s private code to a service.
Python, code indexing, knowledge catalogs
A repository-to-knowledge-catalog tool for compact codebase understanding and fast project orientation.
Why I care about it
Useful when I need to understand a repo quickly or hand better context to an AI tool.
Python, V-JEPA, robotics, planning
A Mac-first robotics scaffold for V-JEPA/V-JEPA2 latent world models, action-conditioned dynamics, planning, and SO-101 adaptation.
Why I care about it
This was me trying to make world-model research less abstract and closer to robot data I could actually inspect.
This is still selective. For every public repo, fork, and unfinished branch, the GitHub shelf has the less edited version.
A shell interceptor and local-AI command agent that catches mistyped commands, suggests corrected terminal actions, and can operate in agent mode from the CLI.
This is the kind of tool I wanted myself: mess up a command, get a better one, and stay in the terminal.
A local macOS assistant that combines voice input, document retrieval, graph-style context, and AppleScript control.
I wanted an assistant that could use my actual computer context without sending everything somewhere else.
A native macOS operations HUD for scripts, launchd agents, Docker containers, SSH services, and remote jobs.
I got tired of checking five places to see what was running, so I put the useful state and controls in one place.
A local-first repository scanner that converts codebases into structured summaries while keeping source private.
I can get oriented in an unfamiliar repo without uploading somebody else’s private code to a service.
A repository-to-knowledge-catalog tool for compact codebase understanding and fast project orientation.
Useful when I need to understand a repo quickly or hand better context to an AI tool.
A Mac-first robotics scaffold for V-JEPA/V-JEPA2 latent world models, action-conditioned dynamics, planning, and SO-101 adaptation.
This was me trying to make world-model research less abstract and closer to robot data I could actually inspect.
An active-inference testbed for backprop-free world models, robotics benchmarks, and smolVLA/LeRobot experiments.
I wanted somewhere concrete to compare what a world model predicts with what a robot actually needs to do.
Aegis Certification Protocol platform for benchmarking robotic policies across safety, mechanical integrity, and mission success.
Robots need more than cool demos, so I worked on clearer pass/fail checks and reports.
A POV video quality pipeline that detects hand presence, stability, lighting, and validity for robot-training data.
Bad training data is expensive. This catches weak trajectories before they reach the model.
A multimodal RAG server with ingestion, indexing, web access, and APIs.
Moved retrieval out of notebook-land and into something an app could actually call.
A personal AI doppelganger experiment trained around Telegram history for reflection and conversational simulation.
A weird but useful experiment in what happens when your chat history becomes something you can talk to.
A Telegram betting and prediction-market bot with TON payments and changing market states.
It forced the chat experience, payments, changing odds, and boring operational details to work as one product.
A Go upload backend using resumable upload hooks and Grably configuration models.
Large uploads fail in annoying ways. This made them resumable and much easier to trust.
A dashboard and upload worker for browsing, previewing, signing, and tracking large file uploads.
Made large upload/storage jobs easier to see, debug, and hand off.
Data exporter and dataset tooling for project configurations, URL retrieval, summaries, and command-driven exports.
Turned messy collected data into exports people could actually use.
A local interview copilot that transcribes conversations and generates real-time suggestions in a web UI.
The interesting part was timing: transcription, context, and suggestions all had to arrive while they were still useful.
A hybrid steganography engine for watermarking images, audio, and text.
Tried different ways to hide and recover watermarks across media without making the media ugly.
A cellular-automata physics engine simulating elemental interactions, heat transfer, and bending-style abilities.
I like systems where a handful of local rules turn into behavior you did not explicitly script.
A tiny hand-crafted decoder-only transformer implementation inspired by Theia Vogel's explainer.
Writing it from scratch made attention, residual streams, and next-token prediction feel real instead of diagram-shaped.
A managed workspace for writing, organizing, and archiving scientific papers.
Built the kind of paper workspace I wanted: drafts, state, and old versions in one place.