I build systems for efficient, reliable LLM inference. My recent work sits at the intersection of model-serving infrastructure, speculative decoding, GPU-aware optimization, and distributed systems.
I am especially interested in the practical systems problems behind modern generation: cache management, scheduling, stateful decoding, observability, and making research ideas hold up in real implementations.
- LLM inference systems — decoding engines, KV-cache management, batching, and serving boundaries.
- Speculative decoding for hybrid models — handling recurrent state alongside full-attention KV caches.
- Distributed systems fundamentals — consensus, fault tolerance, and the infrastructure beneath reliable services.
| Project | What I built | Stack |
|---|---|---|
| hybrid-spec-engine | A lossless speculative-decoding test engine for hybrid-attention LLMs. It snapshots recurrent state, restores it on rejection, and selectively replays only the modules that need recomputation. | Python · PyTorch · CUDA/Triton · Transformers |
| mini-raft-kv | A from-scratch, Raft-backed key-value store for exploring leader election, replication, failover, and state-machine design. | Go · Raft · Distributed Systems |
| Custom-TCP-Stack | A reliable transport-layer protocol built in a custom network simulator, with congestion control and loss recovery. | Python · TCP/IP · Networking |
I study Computer Science at New York University, with a minor in Business Studies. I enjoy turning systems research into clear, testable software and am always happy to connect on LLM serving, inference optimization, and distributed systems.