I'm Philip John Basile, a principal AI systems engineer with 27+ years of software development experience. These are five examples of how I investigate problems, make tradeoffs, and take work through verification.
| Case | What it shows | Evidence |
|---|---|---|
| Fixing quantized matrix multiplication in Apple MLX | Finding a tile-boundary bug, correcting pointer arithmetic, and working through an upstream review | Merged PR, final patch, regression coverage |
| Training and releasing Wisp Coder | Owning tokenizer training, model pretraining, runtime export, and controlled evaluation | Published weights, verification receipts, paired comparisons, and null results |
| Taking training data into model releases | Preserving source records during imports, separating SFT from calibration, and connecting data choices to pruning experiments | Public dataset, verifier and import code, import receipt, pruning report, and paired evaluation |
| Measuring local inference fairly | Separating model speed from streaming behavior and choosing comparable measurements | Pinned models, benchmark protocol, output hashes, recorded results |
| Connecting enterprise systems to AI agents | Owning tool contracts, identity, operational controls, and rollout | Public career case study and documented operating outcomes |
The MLX, Wisp, data, and inference cases link to public code and technical records. The enterprise case summarizes my public portfolio; client source and operational data remain private. Historical measurements are dated and are not presented as new benchmark runs.
GitHub profile · Portfolio and career · Hugging Face · Email