ML engineer focused on LLM post-training, alignment / RL loops, and inference tooling — the glue that turns research ideas into something you can run, measure, and ship.
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; one-factor ablation of RL mechanisms)
- Fast inference paths when serving stacks lag research realities (PEFT / LoRA merge; prefix KV reuse)
- Frontier model mechanisms (MLA, MoE, DSA, DSpark) as readable maps, not vibes
- Practical OSS hygiene — know when a fix is already upstream before opening a PR
Languages I reach for most: Python · C++ · Shell
| Repo | What it is |
|---|---|
| jev-any-llm | Wrap any instruct LLM into Jev-mode prediction — with KV-share branching so shared prefixes reuse one prefill |
| kimi3-align | Open one-factor ablation testbed for K3-style RL mechanisms on AReaL |
| deepseek-mechanism-atlas | Bidirectional wiki + mdBook of DeepSeek V1→V4 mechanisms (MLA / MoE / DSA / DSpark). Docs site |
| 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) |
| 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
- GitHub: @fooSynaptic