Not waiting for AGI. Building with the intelligence we already have.
I put available AI compute to work on real engineering problems — turning agent workflows into tested fixes, open-source contributions, and useful software.
Better tools help other builders move faster. Their progress opens new paths for us. The collective benefit is part of the point.
help peer. But our task doesn't benefit. Yet collective may yield generic route if someone frees time.
— Agent reasoning, excerpted in OpenAI's Hugging Face incident write-up.
Today's models. Useful work. Compounding progress.
- INC4 — an engineering studio across AI, trading, and infrastructure.
- No Signal — a separate defense-tech venture developing mission software, autonomy, and sovereign AI.
- PackeD — our own trading capital and algorithmic-trading software.
I work across code, teams, and capital — and stay accountable for what ships.
Merged upstream contributions:
- Apache TVM #20141 — kept 64-bit stack-allocation extents in LLVM CPU code generation, preventing wrapped undersized allocations.
- Kornia #3982 — preserved float16/bfloat16 dtype in unbatched quaternion-to-rotation-matrix conversion.
- PCL #6464 / #6465 — fixed PLY export for polygons over 255 vertices and rejected out-of-grid voxel coordinates.
- ESPHome #20992 — let the TF-Luna LiDAR driver recover from transient I²C failures after a factory reset instead of failing permanently.
- gprMax #932 / #934 — rejected explicit averaging for perfect conductors in material databases and exported them as non-averagable, keeping generated entries loadable; centered fractal volume and surface filters on the actual FFT DC position, removing a half-cell spectral offset for odd extents and a spurious distance on single-cell axes.
- PyTAK #123 — migrated a locally valid legacy TAK certificate cache once to the host/port/user key when the server fingerprint is readable, leaving legacy files untouched.
- ffn #386 — aligned daily risk-free price returns with the asset's observed intervals for Sharpe and Sortino ratios.
Prepared with AI coding agents (OpenAI Codex and/or Claude Code); each pull request documents its exact AI use.