Navilan Santhanakrishnan

Projects

AI inference

Doubled Qwen3.6-27B decode speed on an Apple M5 Pro, from 14.6 to 29.4 tokens per second, with Metal GPU kernels and speculative decoding in llama.cpp. Every change is measured with a frozen A/B harness and logged, failures included.

C++ · Metal · llama.cpp · GPU kernels · Code · Kernels

Runs DeepSeek V4 Flash on Apple silicon machines without enough memory to hold it, by streaming mixture-of-experts weights from SSD through a slot-bank cache. Fork of the DwarfStar 4 engine.

C · Metal · Apple Neural Engine · MoE · Code

Agents and security

Intercepts any app's HTTP traffic and reverse-engineers it into a typed OpenAPI spec and a runnable MCP server, so AI agents can call the recovered API as a tool.

Rust · Python · TLS proxy · Chromium CDP · MCP · Code

Systems

Runs Windows games on Apple silicon with Wine and D3DMetal or DXMT, choosing the renderer automatically and diagnosing failures.

Rust · C++ · Wine · Metal · Code