Hey, I'm Ankush Kapoor

A manual, from-scratch PyTorch implementation of the 'Attention Is All You Need' Transformer architecture for NMT. Built without high-level abstractions, featuring custom multi-head attention and beam search decoding. Trained on OPUS Books with Noam scheduling and TensorBoard experiment tracking.

A custom 32-bit protected-mode kernel built from scratch in C and NASM, featuring a hand-written bootloader, GDT/IDT setup, ISR/IRQ handling, paging, screen output, and a Round Robin scheduler, all without standard library dependencies.

A 32-bit custom memory allocator that mimics malloc/free over a 1 GB virtual heap, with packed headers, macros for KB/MB/GB allocation, a show() debugger, and a destroy() routine that securely zeroes memory to prevent data leakage.

A multi-source AI research assistant that orchestrates tools with LangGraph, retrieves data from Google, Bing, and Reddit via BrightData, and uses a local Ollama LLM to synthesize context-rich, traceable answers.

An autonomous agent that speedruns Wikipedia by selecting optimal outgoing links using SBERT vector similarity and a Gemini-based reasoning agent, with automated Playwright-driven gameplay, JSON run logging and visual replay.

A Python-based code editor with tabbed editing, syntax highlighting via QScintilla, Jedi-powered autocomplete, and a command runner for building and executing code from within the editor.

A NEAT-powered Flappy Bird implementation where AI agents learn to play by evolving neural networks over generations, visualized in real time using Pygame.