building things in AI agents, inference, analytics, infra, data systems.
previously worked on LLMs, RLHF, small models, and a lot of data pipelines.
- AI agents & orchestration
- Real-time analytics & data infra
- LLM pre-training, fine-tuning, RLHF, small models
- Distributed training & inference at scale
- Systems that just work
- Inference — fast local LLMs on Apple Silicon: Metal kernels and speculative decoding (fastkernel)
- Agent harnesses & multi-agent systems — agents that plan, use tools and work together
- LLM research — process reward models, test-time compute scaling, and code agents for SWE-bench
- fastkernel — the fastest inference engine for Qwen3.8-27B and Qwen3.6-35B-A3B on Apple Silicon, 1.11× faster than Splash 1.3.0 on an M5 Max
- OpenJev — open-weights decision model: typed answers about text, web pages and screenshots in one forward pass, no training per task
- meow-ssh — SSH auth via browser passkeys (Touch ID / Face ID)
- clickhouse-kafka-connect — my fork with multitenancy features before upstream merge
- email: me@abhishek.it
- linkedin: in/abhishekit
- location: remote