AI and machine learning engineer building agentic systems and on-device AI.
- Joined two vLLM Semantic Router workgroups: Router Models & Inference Runtime and Developer Experience & Ecosystem
- Early access to vLLM-SR's Decision 3.0: ported d3 to Core ML (d3-coreml) and wrote Decision 3.0 on Apple Silicon for the vLLM-SR blog
- Fluid Inference was acquired by Levenza
- FluidAudio supports dozens of commercial applications
- 500,000+ monthly model downloads across different Core ML models on FluidInference
- Achieved over 2,000,000 total downloads and peaked 500,000 monthly downloads for Parakeet TDT V3 Core ML
- **1,000+ Twitter followers in 1 month thanks to FluidUse distilled & finetuned models **
- AI Reseacher
- Finetuning PyTorch models to faster frameworks
- Distilled decision models for inference or games like mario
- Real-time audio AI (transcription, speaker diarization, TTS, VAD, etc)
- Inference Engineering, On-Device ML Engineering, Performance Engineering, Agentic Engineering, ML Engineering
AI Work
- FluidAudio — speech AI SDK for iOS and macOS (2.8K+ ⭐)
- FluidUse — Core ML runtimes for decision models in Swift, including Decision 3.0 (
D3Manager) - möbius — model conversion and deployment across accelerators
Some Experiment Highlights (Swift runtimes in FluidUse)
- Decision 3.0 — d3-lite, nano and mini on Core ML; d3-lite triages 1,000 pull requests into vLLM-SR workgroups live on an M5 Pro at 0.12 s per PR, 8.3 PRs/s (demo)
- clef-text-0.6B — distilled Cloudflare's 9B clef-flash into a 0.6B decision model that runs 100% on the Apple Neural Engine; 1,000 support tickets triaged in 57 s on an M5 Pro, 48 ms per ticket, within a few points of the 9B on classification (demo)
- EmbeddingGemma 2 — text encoder 100% on the Apple Neural Engine; sorts 10,080 posts into topics live at ~620/s on a MacBook, any topic splits into subtopics in under a second (demo)
- GLiNER2.5-Decide — Fastino's 340M decision model on Core ML: ~4× faster, ~5× less peak RAM, half the size (demo)
- laya — Convai's typed decision model with 99.5% of ops on the Apple Neural Engine, 3.7 ms per decision on an M5 Pro (demo)
Apps and open-source collaborations
- SGLang — fixed the diffusion LoRA merge cache silently reusing stale merged weights after an adapter is rewritten in place, with regression tests (37K+ ⭐)
- Supertonic — ported Supertonic 3 to Core ML for Apple Neural Engine inference on iOS and macOS (13.7K+ ⭐)
- mlx-audio — added Kokoro-82M v1.1 Mandarin support in MLX (7.9K+ ⭐)
- vLLM Semantic Router — workgroup member; wrote the Decision 3.0 on Apple Silicon blog post (d3-lite, nano and mini on Core ML, matching upstream on JevBench and typed-decisions, 2.5–4× faster per request on a Mac); fixed external RAG result ranking across query windows, secured Playground feedback for read users, and hardened Dashboard API authorization with mutation-boundary rechecks (6K+ ⭐)
- IREE — fixed a global optimization crash and moved compiler passes into the main codegen tree (3.9K+ ⭐)
- vLLM Production Stack — fixed router static service discovery to reject mismatched model list configs at startup instead of failing at request time (2.6K+ ⭐)
- OpenOats — contributed its horizontal wordmark for marketing materials (2.5K+ ⭐)
- Jobpulse — built an AI-powered job search site for beginners that peaked at 10,000 monthly users
Resume · Twitter · LinkedIn · HuggingFace
