I build high-performance AI systems from the ground up, with a focus on LLM inference, GPU acceleration, post-training, and production-grade agentic architectures.

• 🎓 B.Tech AI/ML student at Dayananda Sagar College of Engineering
• 💻 Software Developer Intern at Diarch Group, building production software and AI-driven systems
• ⚡ Focused on LLM inference & GPU systems — CUDA, Triton, KV caching, batching, and performance optimization
• 🧠 Exploring LLM post-training with GRPO, RLHF, and parameter-efficient fine-tuning
• 🤖 Building agentic AI systems using LangGraph, MCP, RAG, and multi-agent architectures
• 🔬 Interested in understanding AI systems below the abstraction layer, from model architectures and kernels to serving infrastructure

A custom Qwen2.5 inference and serving engine built from the ground up, focusing on efficient GPU execution and high-throughput serving.

Highlights

• Implemented custom CUDA/Triton kernels
• Built KV-cache management and dynamic request batching
• Designed an inference pipeline for concurrent LLM workloads
• Achieved 6.34× higher throughput and 2.58× faster processing on an RTX 4050

Python PyTorch CUDA Triton vLLM

Post-trained Qwen3-1.7B using Group Relative Policy Optimization (GRPO) to improve reasoning efficiency.

Results

• Reduced average reasoning trace length by 22.2%
• Reduced average trace length from 579 → 450 tokens
• Maintained 85% GSM8K accuracy

PyTorch Transformers GRPO RLHF

A C++17/CUDA-based language model training system built to understand the internals of modern ML frameworks.

Implemented

• Tokenization
• Automatic differentiation
• Custom attention mechanisms
• GPU memory management
• Neural network training primitives

C++17 CUDA GPU Systems

An AI-powered security auditing platform using 7 specialized agents coordinated through Model Context Protocol (MCP).

The system analyzes source code, cloud configurations, and CI/CD infrastructure through specialized agent workflows.

Python LangGraph MCP Next.js

PyTorch Transformers vLLM GRPO RLHF RAG LangGraph

C++17 CUDA Triton KV Caching Batching GPU Memory Management

LangGraph MCP Multi-Agent Systems Vector Databases Tool Calling

Python TypeScript FastAPI Next.js Docker PostgreSQL MongoDB

I'm interested in collaborating on LLM systems, GPU acceleration, inference optimization, post-training, and agentic AI.

• 💼 LinkedIn
• 🤗 Hugging Face
• 💻 GitHub
• 📧 Email

Building AI systems closer to the metal ⚡