I make LLM architectures from scratch, post-training RL systems, and autonomous agent infrastructure from first principles. Graduating from IIT Madras.

Experience:

• Independent Systems & ML Researcher. September 2024 – Present. Implementing frontier model architectures, test-time search, and RL alignment algorithms from scratch (World Models, GRPO, MoE, custom autograd, LoRA/DPO).
• AI Engineer at LOQO.AI. January 2024 – August 2024. Architected multi-modal video intelligence and automated narrative generation pipelines using Whisper, LangChain, and diffusion models. Reduced manual production turnaround by ~90%.

Education:

• Indian Institute of Technology, Madras (IIT Madras). 2023 – 2026 (Expected).

Open Source Contributions & Merged Pull Requests:

Total Merged Pull Requests: 65 across frontier AI agent runtimes, robotics foundation data pipelines, and developer tooling.

World Models:

• dreamer4-coinrun : Trained a 1.57B-parameter action-conditioned world model with flow matching + shortcut forcing on 9.6M self-generated CoinRun frames on an H200 (PSNR 40.41, beating GenieRedux baselines). Debugged and fixed 5 upstream bugs in the original codebase that prevented environment rollouts.
• world-model-from-scratch : Implementation of Ha & Schmidhuber (2018) World Models. Trains a VAE visual compressor and MDN-RNN memory network (~167k parameters) to learn environment dynamics from raw pixels and simulate game rollouts entirely within hallucinated latent dreams.

LLMs from Scratch:

• llm-lite : Complete, zero-dependency transformer stack in pure PyTorch/NumPy. 8 stages built from scratch: custom BPE tokenizer, autograd engine, AdamW optimizer, RoPE + Multi-Head Attention + KV-Cache, pre-training, SFT, LoRA, and INT4 quantization with comprehensive test coverage.
• nano-gpt-oss : Transformer framework outperforming the GPT-2 baseline across 18 ablation configurations. Features gated Mixture-of-Experts (MoE) routing, SwiGLU activations, Grouped-Query Attention (GQA), and Sliding Window Attention with sink tokens.
• DeepSeek-From-Scratch : Ground-up PyTorch implementation of DeepSeek architecture innovations, including Multi-Head Latent Attention (MLA) with decoupled RoPE compression and DeepSeekMoE fine-grained sparse expert routing.

Post-Training & RL:

• mercor-grpo-agent : Empirical reproduction and critique of Mercor's RL post-training setup vs. standard DeepSeek GRPO on SWE-bench. Investigated length hacking, token entropy regularization, and dynamic DPPO prefix-masking with an interactive benchmark runner.
• swe-in-prod : From-scratch implementation of Group Relative Policy Optimization (GRPO, DeepSeek-R1) for agentic reasoning and code synthesis. Built the actor-environment-reward loop, advantage normalization without a critic network, and LoRA policy optimization on real SWE-bench tasks.
• Reasoning-for-LLMs : Test-time compute scaling, Chain-of-Thought (CoT) search algorithms, and Process Reward Models (PRM) with guided beam search across multi-step mathematical and algorithmic reasoning benchmarks.

Agents in Production:

• Slack-ClawdBot : Production TypeScript agent infrastructure integrating semantic RAG over historical chat threads, persistent memory using mem0, Model Context Protocol (MCP) clients for GitHub & Notion, and 59 custom tool executions. Docker-deployed.
• Vizuara-Agents-10Day-Bootcamp : Multi-agent orchestration architectures spanning LangGraph, CrewAI, and Smolagents for automated code review desks, CI log triage, and deterministic tool-calling workflows.

Technical Focus:

• PyTorch, NumPy, CUDA, Custom Autograd, Flow Matching, World Models
• GRPO, PPO, DPO, Process Reward Models (PRM), LoRA / PEFT, Test-Time Search
• Model Context Protocol (MCP), TypeScript, LangGraph, RAG, Modal GPU Infra, Docker

Contact:

• Email: shobhitagnihotri416@gmail.com
• Portfolio: shobhitagnihotri69.github.io
• GitHub: shobhitagnihotri69