Harshita Mangal
Machine Learning Engineer at Meta
San Francisco Bay Area United States
Summary
🤩
Rockstar🎓
Top SchoolHarshita Mangal is a Machine Learning Engineer with six years of experience building and productionizing quantization and compression solutions for deep learning models, now optimizing inference for Ads models at Meta. She led software-hardware co-design and production rollouts at Qualcomm, where her work on AIMET and LoRA adapter integration contributed to features shipped across millions of chips and enabled low-bit quantization for LLMs in ONNX. Comfortable bridging research and engineering, she translated state-of-the-art techniques—like cross-layer equalization, Adaround, and auto-mixed precision—into robust, framework-agnostic tooling across PyTorch, ONNX, and TensorFlow. A UC San Diego master’s graduate in Electrical Engineering with a computer vision background, she mentors engineers and drives scalable ML design while contributing open-source enhancements (e.g., extending AIMET to support linear and transposed conv layers). Pragmatic and curious, she combines low-level model optimization expertise with system-level deployment experience to accelerate ML inference on real hardware.
6 years of coding experience
8 years of employment as a software developer
BITS Pilani, Birla Institute of Technology and ScienceUniversity of California, San Diego
English, Hindi
Github Skills (13)
net10
quantization10
compression10
pytorch10
quants10
machine-learning10
lossless-compression10
deep-learning10
onnx10
python10
compress10
mask-rcnn9
faster-rcnn9
Programming languages (1)
Github contributions (3)
quic/aimet
May 2020 - Jan 2023
AIMET is a library that provides advanced quantization and compression techniques for trained neural network models.
Role in this project:
ML Engineer
ML EngineerContributions:33 reviews, 99 commits, 234 PRs in 2 years 8 months
Contributions summary:Harshita contributed to the AIMET library by implementing support for linear layers and transposed convolutions, enabling the library to handle more complex neural network architectures. Their work included adding new op types for the conversion of PyTorch types to ONNX, integrating data subsampling logic for weight reconstruction in channel pruning, and adding features for model compression. Additionally, they removed batch norm ops and enabled cross-layer equalization for transposed convolutions.
pytorchtechniquesdeep-learningpruningcompression
quic-mangal/aimet
May 2020 - Feb 2025
AIMET is a library that provides advanced quantization and compression techniques for trained neural network models.
Contributions:8 pushes, 166 branches in 4 years 10 months
pytorchtechniquesdeep-learningcompressionmachine-learning
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.