Final-year B.Tech CSE student at IIIT Kota, working on compilers and machine learning.

I build compilers on LLVM, from language frontends to optimization passes and code generation for CPUs and GPUs, and deep learning systems for medical imaging.

• 🔧 Compilers: wrote minic, a compiler from a C-like language to native executables and GPU assembly, and a set of LLVM optimization passes
• 🧠 Machine learning: first author of a paper on multi-teacher knowledge distillation for chest X-ray diagnosis (under review)
• 🏆 Ranked 1,480 of 32,000+ teams in the Amazon ML Challenge
• 📫 chauhanaditya638@gmail.com

A complete compiler in C++17 on LLVM 18.

flowchart LR
    A[minic source] --> B[Lexer] --> C[Parser] --> D[Sema]
    D --> E[LLVM IR]
    E --> F[Optimizer<br/>-O1 to -O3]
    F --> G[clang + runtime] --> H[x86-64 executable]
    F --> I[GPU backend] --> J[GPU assembly<br/>PTX]
    style H fill:#0e75b6,stroke:#0e75b6,color:#ffffff
    style J fill:#6a3fb5,stroke:#6a3fb5,color:#ffffff

• Hand-written lexer and recursive-descent parser with precedence climbing and error recovery; scoped type checker with Clang-style file:line:col diagnostics
• LLVM IR generation: SSA through alloca + mem2reg, short-circuit &&/|| with phi nodes, guarded division that traps instead of hitting undefined behavior
• -O1 to -O3 through LLVM's new pass manager; native executables linked with clang and a small C runtime
• GPU code generation: emits PTX assembly through LLVM's NVPTX backend (--emit-ptx --gpu-arch=sm_80), optimized with the target's data layout and cost model
• 44 lit/FileCheck tests, including programs run end to end at -O0 and -O2; CI with -Werror, ASan/UBSan, and clang-format

$ minic -O2 examples/collatz.mc -o collatz && ./collatz
111

$ minic --emit-ptx -O2 examples/gpu.mc -o gpu.ptx    # the same language, compiled for a GPU

$ cat bad.mc
fn main() -> int {
  let x: int = 1 + true;
  return x;
}
$ minic bad.mc
bad.mc:2:18: error: invalid operands to binary '+' ('int' and 'bool')
  let x: int = 1 + true;
                 ^

• inst-counter: an analysis pass counting instructions by kind
• my-dce: dead code elimination iterated to a fixed point, keeping stores and side-effecting calls
• unroll-advisor: a loop-unrolling heuristic built on LoopInfo and ScalarEvolution
• lit/FileCheck tests, GitHub Actions CI, and documented design decisions

First author, with Divyansh Sahu, Ayush Singh, and Dr. Veena Khandelwal (IIIT Kota). Under review.

A web application that assists clinical diagnosis with a custom-trained PyTorch model. Built with PyTorch, OpenCV, Streamlit, Supabase, and Plotly.

Ranked 1,480 of 32,000+ teams (top 5%), with an F0.5 score of 97.2%.

Compilers & systems: C++17 · C · LLVM · CMake · Ninja · lit/FileCheck · Linux Machine learning: Python · PyTorch · OpenCV · Streamlit