Akshat Shrivastava is Perceptron AI’s co-founder and chief technology officer. He researches multimodal models and efficient inference, following work at Meta’s FAIR and Reality Labs teams on language, speech and on-device AI.[2][3]
Start with LayerSkip →
Education
Bachelor’s degree, Computer Science; Data Science specialization
Completed 2018
[2]Master’s degree, focus in Machine Learning and Natural Language Processing
Completed 2019
[2]Links
Website
GitHub LinkedIn 
X Co-founder and chief technology officer · Perceptron
Career
Shrivastava earned computer-science and master’s degrees at the University of Washington in 2018 and 2019. At Meta, his work spanned language AI for AR devices, model inference and multimodal pretraining. In November 2024 he and Armen Aghajanyan left FAIR to build Perceptron, focusing on models that process information from the physical world.[2][3][4]
Contributions
Use one model to draft and verify tokens
Shrivastava is an equal-contribution author of LayerSkip. The training method combines layer dropout with an early-exit loss so earlier layers can make useful token predictions. At inference time, those layers draft tokens and later layers verify them. This self-speculative approach reuses the same model’s computation and activations, avoiding the need for a separate draft model.[5]
Generate structured meaning more efficiently
His first-listed-author work on span-pointer networks addresses task-oriented semantic parsing. Instead of generating every output word in sequence, the model predicts a structured representation using pointers to spans of the input. The paper studies the speed and accuracy of this non-autoregressive approach on semantic-parsing benchmarks.[7]
Tools, writing & talks
Tools
Start here
LayerSkipgithub.comCode for early-exit training and self-speculative decoding within one language model.
Start with the early-exit and self-speculative decoding diagram in the linked paper, then the repository’s inference example. Transformer-inference knowledge helps; speedups depend on the checkpoint, workload and exit layer.
Research co-author[6]
Writing, research & talks
- Span Pointer Networksarxiv.org
A non-autoregressive approach to turning utterances into structured intent representations.
First-listed co-author[7]