Shobhit Agnihotri

AI Engineer · IIT Madras

AI Engineer specializing in implementing foundation models from scratch and scaling them for production. I bridge the gap between research and deployment—building LLM architectures (MoE, MLA), optimizing inference pipelines for high-throughput serving, and designing robust, evaluation-driven autonomous agents.

Research Interests

  • LLM architecture — MLA, GQA, Mixture of Experts
  • Inference engineering — KV-cache, quantization, serving
  • RL for code generation — GRPO, reward shaping
  • Autonomous agents — tool use, self-correction
  • Model evaluation & fine-tuning — LoRA, QLoRA
email · github · linkedin · resume

Selected Work

DeepSeek From Scratch PyTorch

Full implementation of DeepSeek architecture — Multi-Head Latent Attention, MoE routing, RoPE, multi-token prediction. Built iteratively over 6 months.

NanoGPT — GQA + MoE PyTorch

GPT implementation with Grouped Query Attention, Mixture of Experts, RMSNorm, and cosine LR scheduling. Trained on TinyStories.

RL Coding Agent GRPO · RL

Autonomous bug-fixing agent trained with GRPO reinforcement learning. Parses issues, executes commands in sandbox, generates unified patches.

AI Agents Framework LangGraph · CrewAI

Multi-framework agent systems — browser automation, agentic RAG, multi-agent collaboration with CrewAI, LangGraph, and Smolagents.

LLM Inference Engine in progress

Production inference pipeline — KV-cache optimization, speculative decoding, continuous batching, INT4/INT8 quantization.

LLM Evaluation Suite in progress

Evaluation framework with custom benchmarks, automated scoring, and comparative analysis across model families.