Shubhamai

Exploring the intersection of low-level systems, hardware, and AI.

GitHub

Projects

  • Great Work

    Great Work

    [WIP] Processing over billions of websites and TBs of data to find the best work using AI.

    The project uses Common Crawl data to process monthly indexes and using 4o-mini to filter millions of sites and perform a certain task.

    Exploring capabilities such as indexing all of the technical blogs on the internet, finding unique early stage startups, etc.

  • Silo

    Silo

    Exploring the Next Generation of Serverless Platforms. Pulling and Launching containers in couple of seconds.

    Built a toy architecture inspired by Modal Labs to pull and run containers in seconds.

    The architecture includes several parts such as a gRPC server, A FUSE filesystem to mount a container filesystem and a custom TCP server to send files saved in a content addressed storage.

  • StableView

    StableView

    Easy, fast and efficient Head Tracking application using only webcam.

    Uses Rust and ONNX to run a realtime inference of an ML model in 30 Hz at very low CPU usage to process webcam images and output head pose.

    The software has over 500 downloads mainly used in flight simulators such as Microsft Flight Simulator, Supports Windows, Mac and Linux.

  • AISandbox

    AISandbox

    Designing AI systems graphically via a node-based editor.

    A node editor built using React Flow to chain and execute multiple AI models together graphically. It also allows to run graphs using API with authentication.

    The web app is built using NextJS, shadcn, Supabase and uses providers like Replicate to run models.

  • n1

    n1

    Simple CPU with custom ISA and assembler in rust.

    This was a very simple 16-bit, single-issue, scalar processor with a Harvard architecture. The CPU is built in Verilog along with an assembler written in Rust.

    It has over 14 different instructions and supports functions (including nested and recursive), loops and branching.

  • grad

    grad

    A compiler for a toy programming language in Rust.

    Converts the source code into an AST and then an executable bytecode, which is then interpreted by a custom stack based virtual machine (VM).

    The parser is a combination of Pratt Parser for expressions and a recursive descent parser for statements.

  • Tsar

    Tsar

    <$80 general purpose flight controller for autonomous vehicles.

    Designed a low cost & bom general purpose flight controller for autonomous vehicles. It is powered by dual core ESP32-S3 and STM32G0 as an IO extender.

    This was the third iteration and has in-built accelerometer, gyroscope, barometer, GPS, radio, SD card, external power IO and over 30 auxiliary IOs.

  • Drone

    Drone

    Drone with custom flight controller, firmware, and ground communication systems.

    A simple hover program built in C uses a Madgwick filter to calculate orientation from sensors which then feeds into a PID controller for hover.

    It also communicates with RPI4 which send/receive data via I2C and websockets. The really cool looking ground control is built using egui in Rust

  • Airliner

    Airliner

    Co-Pilot and Virtual ATC for Microsoft Flight Simulator

    Virtual Air Traffic Controller powered by Whisper, GPT4 and ElevenLabs for simulators and training purposes.

    Implemented with Tauri & Svelte for Desktop Application, Nextjs 13, Supabase.

  • VQGAN

    VQGAN

    Implementation of VQGAN in PyTorch from scratch.

    The project implements Taming Transformers for High-Resolution Image Synthesis in PyTorch from scratch.

    Contains a complete explanation of the architecture with scripts to train and experiment with logging and visualizations.

Miscellaneous

  • Contributions to HuggingFace Transformers. Implemented some papers - VQGAN, Vision Transformers and MobileNet in PyTorch with documentation.
  • Managed a team of 3 in organizing and coordinating 8 different machine learning competitions at AIcrowd, attracting over 2.5k participants, 9k+ submissions, and $4k+ in prizes.
  • Did a course at Udemy on concepts of computer vision and deep learning, with over 100+ lectures and 16+ hours of video content. Rating of 4.5 out of 5 from over 40 reviews.
  • Was ML Teaching Assistant at Zero To Mastery for a while, answering thousands of questions related to ML from learners.
  • Worked as a ML freelancer for many small startups & individuals. Completed over 25 gigs with 12 five-star reviews.
  • 2x Finalist and 1x Sponsored Prize Winner at ETHGlobal.
  • Had First Position at Quantitative analysis of geographical (seismic) data using Machine Learning.
  • Honorable Mentions from Devpost for AISandbox.