PTXAS makes register allocation decisions at compile time. The GPU only reveals what actually worked at runtime. I'm building a system that closes that gap.

Hello I am Sanchit, Final-year Computer Engineering student at VIT Pune (2026). Working at the intersection of compilers, ML and systems and also an athlete.

Currently building "Let It Compile", a PPO-based reinforcement learning agent that observes live CUPTI hardware counters (achieved occupancy, L2 hit rate, DRAM bandwidth, SM utilization) and NVML telemetry (GPU utilization, temperature, power draw) to adaptively select --maxrregcount and block-size configurations that outperform PTXAS static heuristics. The agent treats the compiler as a black box — no PTXAS source access required, optimization happens entirely through its public interface.

Compilation stack: Python → Numba JIT → LLVM IR → PTX → PTXAS → SASS → CUPTI feedback loop
Tech: CUDA, PTX/SASS, CUPTI, NVML, PTXAS, Numba, PyTorch, PPO, Gymnasium, C/C++, Python

Research-driven throughout coursework. Multiple Scopus-indexed publications (IEEE, Springer).
Open to systems research, entry-level roles: Software Engineer / AI Engineer / Researcher.