Full reproduction of the 124M parameter GPT-2 from scratch: built the transformer (blocks, MLP, causal multi-headed self-attention), loaded released weights to verify correctness, then trained from scratch with bfloat16, torch.compile, FlashAttention, DDP across 4 A100 GPUs. Trained on FineWeb-Edu 10B; evaluated on HellaSwag. Our model beats the reference GPT-2 124M eval score.

Worked through Andrej Karpathy's Zero to Hero series, coding each lecture rather than just watching, and documented the journey in a technical blog. Built the language model stack from first principles: autograd engine, bigram character-level model, MLP, batch normalization, WaveNet-inspired model, character-level transformer, BPE tokenizer, and finally GPT-2 pretrained on 4 A100 GPUs.

A Git-like chat session management tool for coding agents. Implementation available as an open-source MCP server, with a modified mini-swe-agent integration for benchmarking.

AI-powered system that tracks fridge inventory using computer vision, reduces food waste, and suggests recipes

Multithreaded Lock-Free Web Server in Rust

Efficient N-body Simulation using the Barnes-Hut Algorithm

Implemented a 2-Phase Commit Simulation in Rust

GPU Accelerated K-means Clustering Algorithm

Engineered a high-performance, multithreaded algorithm for identifying equivalent Binary Search Trees

Implemented Distributed Data Parallel training of VGG11 model on CIFAR-10 dataset using AWS SageMaker, analyzing scaling performance across multiple GPUs to balance parallelism vs. communication overhead in gradient synchronization

Applied LoRA for parameter-efficient fine-tuning of TinyLlama model on the guanaco dataset by injecting LoRA adapter modules into MLP linear layers, testing how training latency was affected by LoRA rank and module size

Quantized facebook/opt-1.3b model to 3-bits while preserving performance by protecting top 1% of salient weights through hardware-friendly scaling approach rather than mixed-precision, with ablation study revealing optimal scaling factors

Implemented a customized RAG model centered on the "Bhagavad Gita," optimizing output by referencing an authoritative knowledge base.