Open to difficult, useful problems
I’m a systems-oriented software engineer working across Python, Linux, containers, application security, applied AI, desktop software, and game systems.
Core Maintainer
Security Manager
Applied AI Builder
Indie Game Developer
01 / Selected evidence
Filter the work by the kind of problem you care about. Turn on evidence mode for the ownership and engineering signal behind each project.
Core maintainer and Security Manager work on a self-hosted AI workspace: authorization, owner isolation, fail-closed behavior, path and token boundaries, model routing, OAuth, regression coverage, and CI.
Owned signal: reconstruct the invariant, trace it through code, state, tests, and authorization boundaries, reproduce the failure, then probe the smallest safe fix.
Resumable review tooling that pins Git ranges, hashes artifacts, treats candidate instructions as hostile data, and confines required execution to secretless, offline probe containers.
Owned signal: evidence-bound findings, bounded subprocesses, stale/duplicate guards, immutable probe inputs, and human-controlled publication.
Mood-aware meeting assistance, audio playback, transcript capture, and Gemini-powered chat with persistent configuration, component previews, smoke tests, and tag-triggered releases.
Owned signal: brought UI, playback, transcription, configuration, packaging, documentation, and Linux/macOS/Windows release paths into one operable desktop product.
An end-to-end platform for dataset validation, PyTorch training, top-k and legality evaluation, inference, local play, and cloud GPU lifecycle orchestration, with a standalone Streamlit arena for playing against swappable model artifacts.
Owned signal: connected data generation, training, evaluation, inference, UI, cloud execution, tests, committed behavior specifications, and a separate CPU demo deployment with draggable-board play.
Tool calling, conditional routing, streaming, checkpoints, Qdrant-backed retrieval, local llama.cpp models, safe knowledge-base operations, and deterministic research flows.
Owned signal: explored agent orchestration as an explicit state machine with local models and controlled tool boundaries rather than a chat wrapper.
A compact Minesweeper implementation for desktop and terminal with a native C core, self-tests, CMake builds, and cross-platform release automation.
Owned signal: evidence that the “C for guts” part is real: deliberately small scope, explicit memory/runtime boundaries, and portable builds.
A strange idea taken through the entire loop: review pull requests from a physical apartment, cash out, buy upgrades, automate the work, and live with what the automation gets wrong.
Owned signal: gameplay systems, diegetic UI, performance work, cross-platform builds, release automation, screenshots, public devlogs, and a playable browser build.
02 / Operating model
The speed comes from reducing uncertainty early, not by pretending it does not exist. Pick a phase to inspect the loop.
question / 01
Map the system and trust boundaries. Reproduce the risky behavior. Build the smallest probe that can disprove the convenient story.
03 / Review instinct
High-risk changes rarely fail only where the patch is loudest. I trace the intended behavior through code, persistence, authorization, execution boundaries, and regression coverage.
Read the ATS-friendly versionState the intended invariant and ownership boundary.
Follow code, state, permissions, persistence, and tests.
Separate the PR regression from an existing defect.
Test the smallest safe fix in an isolated lane.
Beyond the headline
A Debian host, a local speech-to-text tool, an ML experiment pipeline, a native desktop app, and a Godot prototype all ask the same question: can someone understand, operate, and continue this after the exciting first pass?
Lab bench
Digital logic, a working 4-bit CPU and custom assembly-like instruction set in Turing Complete, plus IC10 automation. Useful proof of curiosity; not presented as professional processor design.
04 / Contact
Systems engineering, maintainer tooling, security-sensitive application work, applied AI, or an odd prototype that deserves to become real.
