Computer Science undergraduate at IIT Hyderabad (B.Tech CSE, 2024-2028). I like systems, compilers, computer architecture and ML inference: building things close to the machine and measuring whether they actually work.
- Fixing an X86 backend crash in LLVM on vector
atomicrmw(issue #226002, PR #230788, open) - Compiler construction: DSL front ends and lowering to LLVM/MLIR
- Research on speculative decoding and draft-model distillation (write-up on Medium)
| Project | What it is | Stack |
|---|---|---|
| LLVM contribution | Root-caused and fixed an X86 ISel crash: vector atomicrmw operands are now expanded to a cmpxchg loop, with regression tests with and without +cx16 |
C++, LLVM |
| RISC-V Toolchain & Web Simulator | Assembler, disassembler, debugger, cache simulator and pipeline hazard detector; compiles to WebAssembly for a browser front end | C++17, WASM |
| xv6 Kernel Extensions | Syscall-aware MLFQ scheduler, clock page replacement with swap, FCFS/SSTF disk scheduling, RAID 0/1/5-backed swap | C, RISC-V |
| DistillSpec research | Multi-GPU platform for speculative-decoding research: 500+ tracked experiments, 71 checkpoints, multi-rollout distillation that beat the JSD baseline, and delayed tree branching extended to trained drafts (write-up) | Python, PyTorch, vLLM |
| OCaml compiler extension | Exception handling across lexer, parser, AST, typechecker and runtime | OCaml |
| FPU-assisted integer execution | Splitting integer multiply chains across integer and FP units by a throughput model; measured on gem5 (X86 O3) across optimisation levels, IU/FPU latencies and IU counts: about 2x on convolution at -O3 (via compiler vectorisation), no gain on factorial or matmul | C |
| Data Structures | Balanced trees (AVL, red-black, 2-4, B-tree), hashing, graphs and 2-SAT, fuzz-tested against std::set |
C++ |
| Business entity resolution | Record-linkage pipeline for the Amazon ML Challenge: blocking, pruning and a two-pass LightGBM matcher | Python |
| Tune DSL | Music DSL compiled to audio with three cross-checked backends (Python, C++, MLIR) | Python, MLIR |
Operating systems · Compilers · Computer architecture · ML systems · Performance engineering