Projects

End-to-end LLM system with distributed FSDP/DDP training, Flash Attention v2, and Mixture-of-Experts support. Includes Mixture-of-Refusals (MoR), a novel safety routing mechanism achieving 2–3× speedup on safe queries with identical safety guarantees. vLLM inference engine with quantization, speculative decoding, and prefix caching, deployed on Kubernetes with Prometheus monitoring.

Production-ready framework for distributed ML training across SSH clusters, Modal Cloud GPUs, and local multi-GPU setups. PyTorch DDP with automatic gradient synchronization, unified CLI abstracting the complexity of distributed workflows.

Deep learning framework for PDF malware detection with 99.5% accuracy, robust against FGSM and PGD adversarial attacks. Custom malware injection pipeline for training; explainability via Kolmogorov-Arnold Networks. Published in Applied Soft Computing (Q1 Journal).

Qwen2.5-7B fine-tuned on Indian legal datasets using LoRA + PEFT + Unsloth. Deployed with 4-bit quantization and Flash Attention 2 on Modal. Specialized for Indian legal document analysis and question-answering.

Code optimization system using Chain-of-Thought and ReAct reasoning with local LLMs. Autonomous refactoring, syntax analysis, and automated test generation, fully local, no API costs.

Interactive visualizer that animates transformer attention weights in real time as a local LLM generates text. Every new token shows which prior tokens the model attended to, across all 28 layers and 12 heads. Live heatmap, per-layer scrubber with entropy sparkline, and per-head selection. No cloud API required.

Fully decentralized, content-addressable file storage system in Go. Handles streaming of large files across distributed nodes with fault tolerance through decentralized architecture and high-performance concurrent operations.

Distributed 3D printer management system using the Raft Consensus Algorithm for data persistence — replacing traditional centralized databases with a consensus-based distributed log.

Billion-scale vector embedding benchmarks (768, 1024, 2048 dimensions) built from open-source codebases using open-source models. Addresses critical limitations of existing ANN benchmarks, enabling robust evaluation under realistic workloads. Selected for GSoC 2025 (acceptance rate <10%).

Complete AlphaZero implementation from scratch — self-play training, neural network-guided Monte Carlo Tree Search, achieving superhuman board game performance.