Skip to work

STATUS: BUILDING · OPEN TO REMOTE ROLES

Abhishek Ghosh

I build things that go fast: CUDA kernels, Rust backends, AI agents, whole products.
From a thread on the GPU to the screen a user taps.

See my GitHub Field notes
SPECIMEN 001 · A. GHOSHink on dust
SCROLL
02 / FIELD SPECIMENS 5 PLATES · SECTOR 7

Low-level when it matters, full-stack when it ships.

Five specimens from the bench, catalogued from the metal up: kernels written by hand, tools that make the GPU legible, and systems built to ship.

SPECIMEN 01 · SECTOR 7ACTIVE · NOW

gpu-kernels

Hand-written CUDA C++ from first principles, profiled with Nsight Compute and measured against cuBLAS.

  • CUDA C++
  • Nsight Compute
  • cuBLAS baseline
  1. K.01online softmax
  2. K.02reductions
  3. K.03tiled GEMM
  4. K.04fused attention
SPECIMEN 02 · SECTOR 7PL. II

cuda-visualizer

See how threads, warps and blocks map onto your data.

  • CUDA
  • Threads · warps · blocks
SPECIMEN 03 · SECTOR 7ACTIVE

kernel-crucible

An RL environment where an agent writes CUDA kernels and is rewarded for correctness and speed.

  • Reinforcement learning
  • CUDA
  • Reward design
SPECIMEN 04 · SECTOR 7ACTIVE

placnoix

Modular university ERP/LMS: attendance, gradebook, assessments, HR, finance. Backend moving to Rust.

  • ERP / LMS
  • Rust backend
  • Modular
SPECIMEN 05 · SECTOR 7PL. V

strata-vision

Hierarchical fine-grained vision engine, from the whole object down to the detail that separates two classes.

  • Computer vision
  • Fine-grained
  • Hierarchical

03 / INSTRUMENTS

Tools I reach for, and why.

FIG. 03 · bench inventory · each one earns its place

  1. 01 CUDA & C++when every microsecond counts kernels
  2. 02 Rustbackends that don't fall over axum · tokio
  3. 03 Python & PyTorchmodels, RL, experiments research
  4. 04 TypeScript & Next.jsthe part users see interfaces
  5. 05 Postgres, Docker, Linuxkeeping it all running ops
FIG. 03a · ASTROLABE θ 000.0°
three orbits · twelve instruments · one core

04 / TRAIL

How I got down the stack.

Route log · four waypoints · ink follows your scroll

  1. 2024

    WP-01 · km 0

    Started B.Tech CSE at LPU

    Computer science at Lovely Professional University, class of 2028. The trail starts here.

  2. EARLY

    WP-02 · the surface

    Full-stack & Web3 products

    Whole products, end to end: TypeScript and Next.js up front, Postgres behind, Solana for the on-chain parts.

  3. SIH ×2

    WP-03 · under a clock

    Hackathons

    2× Smart India Hackathon internal qualifier at college level. Building fast, with a team, against a deadline.

  4. NOW

    WP-04 · you are here

    Down the stack

    Rust systems and hand-written GPU kernels: tiled GEMM, fused attention, measured against cuBLAS.

05 / CURRENTLY

On the bench right now.

field-terminal · tty0

whoami: abhishek, systems engineer: CUDA · Rust · AI

now: gpu-kernels: tiled GEMM → fused attention. placnoix: backend → Rust. kernel-crucible: rewards for faster kernels.

looking for: remote GPU, inference and systems roles

END OF TRANSMISSION · REPLY ON THE WIRE

Let's build something fast.

github.com/Codenuclei→