Models that do your job, not every job.
When prompting and retrieval run out, the model has to change. I fine-tune open-weight models on your data, distill large models into small ones you can run yourself, use reinforcement learning to sharpen them for a specific language or stack, and train from scratch when nothing existing fits.
What you get
- 01 A trained model and the pipeline that produced it
- 02 An eval suite built from your real cases
- 03 A model card: what it does, where it fails, what it costs to run
- 04 Handover so your team can retrain it
Proof
- Python private
compass model
In progress. Qwen fine-tuned for terminal and agent work, distilled and RL-tuned into 4.5B variants per stack, trained on one RTX 4090.
- Python code ↗
lora
Fine-tuning a reasoning model, small and effective.
github.com/jtrefon/lora
- Python code ↗
llm-mini
A transformer LLM built from scratch with minimal dependencies.
github.com/jtrefon/llm-mini
- Rust code ↗
blt
A high-performance byte-level tokenizer in Rust.
github.com/jtrefon/blt
- Rust code ↗
distiller
Model distillation in Rust.
github.com/jtrefon/distiller
Lab notes on this
- amber-bench Local models for agentic coding, measured.
- 2026-10-27 Attention Is All You Need: the paper, the idea, and what it changed
- 2026-10-27 Decoder-only transformers: the architecture behind every chat model
- 2026-10-27 DeepSeek: the architecture behind V3 and R1, and the DeepSeek moment
- 2026-10-27 Dense vs mixture-of-experts: what you trade, from routing to serving
- 2026-10-27 Diffusion models: from noise to images, and now to text
- 2026-10-27 Distillation: logit KL, temperature, and on-policy vs off-policy
Scope check
Don’t hire me for this if
A better prompt and good retrieval would solve it. I’ll check that first.
Have a problem like this?
Describe your problem30-minute technical call with the engineer who builds it. No sales team, no slides.