Skip to content
View ubermenchh's full-sized avatar
🎯
Locked In
🎯
Locked In

Block or report ubermenchh

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ubermenchh/README.md

avatar

Umang Kaushik

AI Engineer · ML Researcher


Low-level systems and machine intelligence: inference engines, GPU kernels, small models.

Site  Blog  Resume  X  Hugging Face  Email

Projects

Project What it is Stack             Year
mini‑vllm PagedAttention with Triton kernels, prefix caching and continuous batching Python CUDA 2025
gemm‑optimization SGEMM on an A100: 0.29 → 9.6 TFLOPS, against cuBLAS CUDA C++ 2025
flash‑attention FlashAttention 2 from scratch in CUDA CUDA C++ 2024
rl‑wordle SFT and RL on Qwen3-1.7B. Wins on held-out words: 0% to 13% Python PyTorch Hugging Face 2026
nanohunch Qwen3-4B with LoRA that answers in calibrated probabilities, not text Python MLX 2026
anki‑cli Anki from the terminal, for humans and agents Python SQLite 2026

More: engram · mhc · tangent · void · Flash · pycc · clox

Tech stack

Area Tools
ML PyTorch Hugging Face CUDA MLX
Languages Python C C++ TypeScript Shell Rust Zig Go
Backend FastAPI GraphQL React RabbitMQ Redis Kafka Nginx
Infra & data Docker GitHub Actions AWS Linux Postgres Elasticsearch MongoDB SQLite

Also: JAX · Triton · vLLM · LangGraph · Google ADK · Faiss

Pinned Loading

  1. anki-cli anki-cli Public

    Manage and study Anki flashcards from the terminal. AnkiConnect or direct SQLite backends, Anki search syntax, JSON output for scripts and AI agents, and a TUI for browsing and review.

    Python 62 5

  2. mini-vllm mini-vllm Public

    A minimal, readable vLLM: PagedAttention, continuous batching, chunked prefill and prefix caching in ~900 lines of PyTorch + Triton.

    Python 23 2

  3. engram engram Public

    Python

  4. flash-attention flash-attention Public

    flashy attention

    Cuda

  5. nanohunch nanohunch Public

    Calibrated decision probabilities from a 4B language model with a LoRA adapter (MLX). Includes the experiments, ablations and results on OpenDecisions.

    Python

  6. rl-wordle rl-wordle Public

    Train Qwen3-1.7B to play Wordle with SFT + RL (verifiers taskset, prime-rl, Modal)

    Python