Skip to content
View aryanyadav0402's full-sized avatar

Block or report aryanyadav0402

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
aryanyadav0402/README.md

Hi, I'm Aryan Yadav 👋

AI / LLM Engineer — training, aligning & serving large language models in production.
Incoming MSc Artificial Intelligence, University of Edinburgh · Sep 2026.

LinkedIn


🧭 About

I spent ~3 years at Otsuka Shokai (Japan) building LLM systems end-to-end — from post-training and alignment to multi-node inference serving and shipping the apps on top. I care about making models train faster, align better, and serve reliably at scale.

  • 🔭 Now: deep open-source contributions to the LLM training/serving stack (vLLM, verl, SGLang, TRL).
  • 🎓 Next: MSc AI @ Edinburgh — dissertation extending my RL-for-retrieval work.
  • 🎯 Goal: Research / ML / Applied Engineer roles at frontier AI labs.
  • 💬 Ask me about: RLHF pipelines (DPO/ORPO/PPO), distributed FSDP training, and high-throughput vLLM serving on H100/H200.

🛠️ What I work on

Post-training & alignment — SFT and preference optimization (DPO / ORPO / SimPO / PPO) on models up to 72B, using Axolotl & LLaMA-Factory. RL for agents — applied & benchmarked the s3 RL-search framework on a Japanese domain corpus; debugged the distributed FSDP → vLLM training stack to get it running (precision 47.5% vs 32.8% naive RAG, +14.7 pts on a 50-sample internal eval). Inference at scale — multi-node vLLM serving on H100 / H200; RAG pipelines with rerankers. Shipped products — solo-built an internal pre-visit assistant (MCP server + Databricks + LLM + React UI + CI/CD) and an AI counselor app, plus eval dashboards and data pipelines.


🧰 Tech

Python PyTorch Hugging Face vLLM CUDA Docker React

Post-training: TRL · Axolotl · LLaMA-Factory · verl  |  Serving: vLLM · SGLang  |  Infra: FSDP · Ray · Databricks


Open to research & applied ML engineering roles (UK, 2027).

Popular repositories Loading

  1. Test Test Public

    Jupyter Notebook 1

  2. Dorami Dorami Public

    Jupyter Notebook 1

  3. final final Public

  4. aryanyadav0402 aryanyadav0402 Public

  5. sglang sglang Public

    Forked from sgl-project/sglang

    SGLang is a high-performance serving framework for large language models and multimodal models.

    Python

  6. vllm vllm Public

    Forked from vllm-project/vllm

    A high-throughput and memory-efficient inference and serving engine for LLMs

    Python