I'm an AI engineer and ML researcher in Vancouver. I train small models and ship the products around them — as one person, end to end: research question, training runs, evaluation, product design, backend, infrastructure, releases, and support.
Current focus: sub-100M-parameter models that run locally inside my own products, held to preregistered, blinded evaluation against strong baselines. I also contribute upstream to MLX, Apple's machine-learning framework for Apple silicon.
Tab Namer 77M names PorkiCoder terminal tabs from a short task description. Tiny 12.7M and 35M from-scratch models could pick a real noun and still scramble English. This file starts from Google FLAN-T5-small, then a few hours of extra training on DigitalOcean Ada GPUs.
Weights: huggingface.co/porkr/porkicoder-tab-namer-77m. On 500 held-out tasks the trained model averaged 7.28 against Google's unchanged 3.12 (+4.17). Two later 1,000-task holdouts sit at 7.59 and 7.55.
| Link | What |
|---|---|
| Open weights | 77M Title-SFT FLAN, raw greedy titles, 20 Google-vs-ours examples on the card |
| Session 11 ship note | Continue-SFT on the leak-checked mix; holdout 1000 / 1000b |
| Research index | Earlier campaign notes |
Each of these is designed, built, deployed, and supported solo — models, product, backend, billing, releases.
| Operation | Class | Brief |
|---|---|---|
| porkicoder | Agentic coding assistant | Desktop agent that reads code, edits files, runs commands, and coordinates parallel sub-agents. Ships the Title-SFT FLAN tab namer on-device. |
| resumehog | Resume optimization | Tailors a resume to a specific role by matching what hiring teams actually screen for. |
| hogmatix | X automation | Schedules posts and operates multiple X accounts from one place. |
Patches sent upstream to MLX. The count below is live from the GitHub API.
