Mayank Mishra
Co-Founder at Stealth AI Startup
Berkeley, California, United States
Summary
🤩
Rockstar🎓
Top SchoolMayank Mishra is a research-driven AI engineer and co-founder based in Berkeley with eight years of experience building and optimizing large-scale ML systems. He led pretraining efforts at the MIT-IBM Watson AI Lab and worked at IBM Research on efficient LLM training for supercomputers, bringing deep expertise in distributed training, mixed-precision, and inference deployment. An active open-source contributor, he has made substantive fixes and features in flagship projects like PyTorch and Hugging Face Transformers—improving DDP/FSDP parity, FP16/FP8 stability, tokenizer support for Llama/CodeLlama, and PEFT mixed-precision adaptations. Now pursuing a PhD at UC Berkeley while co-founding a stealth AGI startup, he bridges cutting-edge research with production-grade engineering. Avid reader across quantum computing, information theory, and RL, he pairs broad theoretical curiosity with hands-on contributions to high-impact ML tooling.
8 years of coding experience
5 years of employment as a software developer
Indian Institute of Technology Delhi (IIT Delhi)Doctor of Philosophy - PhD, Computer Science, Doctor of Philosophy - PhD, Computer Science at University of California, Berkeley
Hindi, English
Github Skills (31)
transformers10
tokenize10
pytorch10
distributed-training10
language-model10
python10
pre-trained-model10
machine-learning10
llm10
huggingface-transformers10
tokenizer10
deepspeed10
deep-learning10
natural-language-processing10
transformer10
Programming languages (6)
Github contributions (5)
huggingface/transformers-bloom-inference
Sep 2022 - Jan 2023
Fast Inference Solutions for BLOOM
Role in this project:
Back-end Engineer & DevOps Engineer
Back-end Engineer & DevOps EngineerContributions:8 reviews, 22 commits, 55 PRs in 4 months
Contributions summary:Mayank's commits primarily focused on building and deploying the inference server for the BLOOM model. They moved the server solution and implemented changes to the server and related utilities. This involved setting up the deployment framework using DeepSpeed and gRPC, as well as integrating with the Hugging Face Accelerate library. Mayank's contributions enabled fast inference solutions for the BLOOM model.
nlppytorchbloominferencehuggingface-transformers
huggingface/transformers
Jan 2021 - Feb 2025
🤗 Transformers: State-of-the-art Machine Learning for Pytorch, TensorFlow, and JAX.
Role in this project:ML Engineer
ML EngineerContributions:8 reviews, 16 PRs, 45 comments in 4 years 1 month
Contributions summary:Mayank contributed to the `huggingface/transformers` repository, focusing on implementing and fixing aspects of tokenizers within various models, specifically related to the Llama and CodeLlama architectures. They also added MLP bias configurations, demonstrating involvement in model configuration and potential optimization. Furthermore, their work involved modifications for Granite and GraniteMoE, including refactoring, adding new functionalities, and incorporating shared experts.
pythonbertspeech-recognitionstate-of-the-artflax
Find and Hire Top DevelopersWe’ve analyzed the programming source code of over 60 million software developers on GitHub and scored them by 50,000 skills. Sign-up on Prog,AI to search for software developers.