rusheel = {
    "focus"     : ["ML research", "medical imaging", "oncology", "time series", "LLMs"],
    "currently" : "OSS contributor + Researching + drifting through college",
    "loves"     : ["Anime", "Chess ♟️", "Formula 1🏎️", "Thriller Novels", "Gym", "Taekwondo 🥋"],
    "chess"     : "RedEgnival @ lichess.org",
    "pronouns"  : "he/him",
    "fun fact"  : "Black Belt in Taekwondo",
}

Spinal Cord Shift Prediction — 3D CNN-LSTM built with Nanavati Hospital to segment and quantify spinal cord shifts during radiotherapy using real HNC patient volumetric data.

dicom-meta-lite — lightweight Python package for rapid DICOM metadata extraction, built for medical imaging and oncology research workflows. Available on PyPI.

Chaos Economy - AMD Hackathon GRAND PRIZE, Multi-agent RL sim: 4 traders, market maker & SEC regulator trained via GRPO on Llama-3.2. 250 steps of emergent financial crisis — slaughter, adaptation, collusion, regulatory oversight — no scripts, just learned behavior. Our 3B LoRA (250 training steps) beats Nemotron 120B on PnL (+13.68 vs +7.18) and beats every other model tested. The MI300X's 192GB HBM3 made full BF16 training possible without quantization — no dequantization overhead, no 4-bit gradient degradation. That precision advantage is why a 3B trained on AMD outperforms 7B, 8B, 30B, and even 120B baselines running on NVIDIA-optimized inference.

• 🔬 ML research — especially for medical imaging, oncology & real-world healthcare impact
• 📈 Financial MArkets — options flow, expiry dynamics, institutional analysis
• 🤖 LLMs, TRL, deep learning & AI agents
• ♟️ chess addict — @RedEgnival on lichess
• 🏎️ Formula 1 fan, also dipping my toes into MotoGP
• 🥋 Black Belt Taekwondo
• 🎌 everything anime, manga and light novels — sometimes even real books :)
• so systematic I even track it all: MyAnimeList