I move between research and production: reading a paper on Monday, reimplementing its core idea by Wednesday, and shipping it as a real pipeline by the weekend. Most of my work lives at the intersection of LLMs/SLMs, agents, and real-time multimodal systems (voice + avatars) — usually with a strong bias toward running fully offline/on-device rather than behind an API.
- 🔭 Currently: distilling large-teacher signal into small, deployable models, and building local-first agent/voice tooling
- 🧪 Research habit: pick an industry blog post or paper, reimplement it from scratch, and see where the real bottlenecks are
- ⚡ Philosophy: prototype in Python, push the hot path to Rust or C++ once it's worth optimizing
- 🌱 Currently deep in: agent memory architectures, SLM distillation, and low-latency inference
Languages
AI / ML tooling
Cloud & infra
Where I actually write code
| Project | What it is |
|---|---|
| slm-recommendations | Reimplementation of a DoorDash search-ads paper: an LLM teacher labels query-item pairs, and a DistilBERT bi-encoder student learns to reproduce that relevance signal at a fraction of the cost |
| Local Murmur | Fully local, offline voice dictation for Apple Silicon — hold a hotkey, speak, get text pasted anywhere. whisper.cpp + Metal, zero cloud calls |
| Forge_Flow / agentforge | A platform for building, orchestrating, and deploying agents on local models — automation without a per-call API bill |
| RUST-AGENT-MEMORY-MVP | Small Rust testbed for probing how different memory strategies change agent behavior over long horizons |
| Avatar-research | Research track for real-time talking-avatar and character-animation pipelines |
| SRE-genai-toolkit | GenAI tooling built for SRE workflows |
The graph above blasts contribution cells instead of painting squares — powered by pacman-contribution-graph, refreshed daily by a GitHub Action.