starone01.me · linkedin · x @iamstarone01 · ping@starone01.me
final-year EEE @ Coimbatore · self-taught on a phone through Termux · now deep in LLVM / MLIR
// pipeline.mlir — how I think about building
func.func @starone01(%idea: !prod.prototype) -> !prod.shipped {
%ast = "frontend.parse"(%idea) : (!prod.prototype) -> !ir.ast
%lowered = "midend.lower"(%ast) : (!ir.ast) -> !mlir.module
%opt = "midend.canonicalize"(%lowered): (!mlir.module) -> !mlir.module
%bin = "backend.emit"(%opt) {target = "prod", opt = "-O3"} : (!mlir.module) -> !prod.shipped
return %bin : !prod.shipped
} // scroll ↓ to execute each passI'm Prashanth, known online as StarOne01. Final-year Electrical and Electronics Engineering student, but most of what I do lives in the gap between hardware and software: compilers, model inference, and the systems that make AI run efficiently on real silicon.
Self-taught, starting on a phone through Termux before I ever owned a laptop. Since then: a production multilingual clinical AI system, 16 LLVM contributions, a film engine on custom embeddings, and multilingual LLM research. Now going deep on compiler internals and MLIR, the layer where models actually meet hardware.
I wrote my first production code on a phone. No laptop. Just Termux, a cracked screen, and WiFi that dropped every twenty minutes. People said wait for the right setup. I didn't.
| project | what it is | |
|---|---|---|
op @medclara |
Medclara — Clinical AI · Founding Engineer | Voice-first multilingual clinical docs. IndicConformer ASR + fine-tuned Gemma for SOAP notes. Architecture, infra, and compliance, solo. |
op @movieslikethis |
MoviesLikeThis — Embeddings | The films that stay with you. Same feeling, different film, matched by emotional experience, not genre. Custom embeddings + content fingerprinting. |
op @sherlock_sft |
Sherlock SFT — Fine-Tuning | Character-consistent LM via SFT. Iterated on data quality and debugged a QLoRA NaN loss down to root cause. |
op @bfloat16 |
bfloat16 — Numerics | Lightweight C++ impl of Google Brain's bfloat16. Same 8-bit exponent as FP32, 7-bit mantissa, the format ML hardware actually speaks. |
op @phrasenux |
PhraseNuX — where it started | C++ CLI password manager, AES encryption, zero deps, written entirely on a phone. |
16 contributions across the stack, now pushing into MLIR:
- Backend / codegen: X86 target-specific lowering for
_Float16ops —FABS,FNEG,FCOPYSIGNat instruction-selection level. - Frontend: New Clang diagnostic in the Sema layer for function-like macro references requiring parentheses — 49 review comments, full CI.
- libc docs: Series of POSIX header documentation additions (
dirent,unistd,termios,sys/stat,sys/wait,netinet/in…), each through review.
Started with docs, proved out backend and frontend, now going deeper into MLIR. Progression over assertion.
→ All of it lives at github.com/StarOne01
Languages → Go · Python · C++ ★ · TypeScript / Next.js
Systems → LLVM · MLIR (in progress) · Compilers · System Design · Local LLM Inference · CUDA (in progress)
ML / AI → SFT / QLoRA · ASR Pipelines · Embeddings · RAG
Infra & Data → AWS · GCP · PostgreSQL · Qdrant · Fedora
- Deepening into MLIR dialects and passes — building on the X86 backend + Clang Sema foundation
- Local LLM inference — deploying open models in production, Ollama internals, scheduling engines
starone01.me · linkedin · x · ping@starone01.me