ML engineer working on LLM post-training and inference / serving infrastructure — PEFT adapters, rollout-style training loops, and the glue that makes research code shippable.
I care about the assumptions under the code: simulation and tuning help, but they don't replace understanding the update rule.
- Post-training & optimization loops (GRPO / SAO-style asynchronous rollouts)
- Inference tooling when serving stacks lag PEFT / LoRA realities
- Frontier model mechanisms (MLA, MoE, DSA, DSpark) as readable maps, not vibes
- Practical OSS contribution hygiene (know when a fix is already upstream)
Languages I reach for most: Python · C++ · Shell
| Repo | What it is |
|---|---|
| peft-adapter-merger | Offline PEFT adapter → dense merge, for reliable inference when serving frameworks trail latest LoRA / PEFT |
| Single-rollout-async-Optimization | Unofficial re-implementation of SAO (Single-rollout Asynchronous Optimization) |
| deepseek-mechanism-atlas | Bidirectional wiki + mdBook of DeepSeek V1→V4 mechanisms (MLA / MoE / DSA / DSpark). Docs site |
| upstream-fix-gate | GO/STOP gate before opening OSS PRs — detect fixes already shipped upstream via gh |
| py_pcalg | Lightweight PC algorithm for causal skeleton discovery — pip + CLI + graph viz |
Older course / trial repos stay in the account for history; they are not what I am building toward now.
fooSynaptic — foo as in foo / bar: learning means solving for the unknown; Synaptic for intelligence wired like synapses.
- Blog: foosynaptic.github.io