edge.ys

PhD Scholar · Computer Science & Engineering · IIT Gandhinagar

Making inference cheap enough to run anywhere it needs to.

I'm Yash Sahu. My research finds the best place — on-device, at the edge, or in the cloud — for a neural network to run, and adapts that decision as conditions change, so LLMs and SLMs stay fast and usable on resource-constrained hardware.

Email Projects Research focus
device edge node cloud split point (adaptive)

The split point moves between device, edge, and cloud based on compute load, network cost, and model characteristics — decided by a learned policy, not a fixed rule.

About

Portrait of Yash Sahu

I'm a PhD scholar in the Department of Computer Science and Engineering at IIT Gandhinagar, advised by Prof. Anirban Dasgupta. My work sits at the intersection of systems and machine learning: I study how to run AI inference efficiently when compute, memory, and network conditions can't be taken for granted.

Concretely, that means adaptive split computing — dynamically choosing where a model's computation should happen across a device, one or more edge nodes, and the cloud — and efficient LLM/SLM deployment through quantization, LoRA, and knowledge distillation. I use reinforcement learning (PPO) and surrogate cost models to make these decisions in real time, as system conditions shift.

Before the PhD, I completed my B.Tech in Computer Science and Engineering (Cyber Security) at RGPV, where I built a foundation in data structures, operating systems, computer architecture, and security.

Split Computing Edge AI LLM / SLM Inference Reinforcement Learning Model Optimization

Research & education

  • 2024 — Present

    PhD in Computer Science and Engineering

    IIT Gandhinagar · Adaptive & Resource-Efficient Edge AI · Advisor: Prof. Anirban Dasgupta

    • Adaptive split computing for AI inference across resource-constrained devices, edge servers, and the cloud.
    • Dynamic model partitioning, where the split point adapts to computation, communication, and model characteristics.
    • Multi-hop split computing — hop selection, split-point selection, and collaborative execution across edge nodes.
    • RL-based (PPO) adaptive inference decisions under changing system conditions, guided by surrogate cost models.
    • Efficient LLM/SLM inference: quantization, LoRA, knowledge distillation, and resource-aware deployment using the Hugging Face ecosystem (Transformers, PEFT, Accelerate, Datasets, BitsAndBytes).
  • 2020 — 2024

    B.Tech in Computer Science and Engineering (Cyber Security)

    RGPV University, Bhopal

    • Coursework: Data Structures & Algorithms, Operating Systems, Computer Architecture, DBMS, Machine Learning, Computer Networks, Cyber Security.

Selected projects

Adaptive Split Computing using Reinforcement Learning

Python · PyTorch · PPO

An RL-based inference optimization framework that dynamically selects model split points from system state and predicted inference cost — policy and value networks, PPO training, GAE, clipped objectives, and state-dependent action selection.

Efficient LLM/SLM Inference for Edge Devices

Python · PyTorch · Hugging Face

Quantization, LoRA/PEFT, and knowledge distillation for lightweight deployment, with a focus on how partitioning, communication overhead, and compute cost trade off against inference quality.

VakyaVed — Multilingual Speech Recognition

Python · PyTorch · ASR

Resource-efficient automatic speech recognition for Hindi, Gujarati, and English, exploring Conformer-based architectures, multilingual and transfer learning, and audio preprocessing pipelines.

Rocket Machine Log Anomaly Detection

Python · LSTM · Time-Series

Anomaly detection and summarization over machine telemetry logs for resource-constrained environments, with sensor correlation and grouping strategies for efficient detection.

Graphi — Interactive Graph Editor

Java

A graph visualization and editing application supporting graph construction and algorithm visualization, including Dijkstra's shortest path algorithm.

Turing Machine Simulator

Java

A graphical simulator for designing and executing Turing machines, including state transitions and tape operations.

Technical foundation

Programming Python, Java, C/C++
Machine Learning Machine Learning, Deep Learning, Reinforcement Learning, Neural Networks, Time-Series Modeling, Model Optimization
AI / LLM Large & Small Language Models, Transformers, LLM Inference, Quantization, LoRA, PEFT, Knowledge Distillation, Adaptive Inference
AI Systems Split Computing, Edge AI, Distributed Inference, Model Partitioning, Resource-Aware Computing, Multi-Hop Inference
Frameworks PyTorch, Hugging Face Transformers, PEFT, Accelerate, Datasets, BitsAndBytes, scikit-learn
Tools Git, GitHub, Linux, Jupyter, Google Colab, LaTeX, Overleaf
CS Foundation Data Structures & Algorithms, Operating Systems, Computer Architecture, Cyber Security

Teaching & academic experience

  • Instructor

    CohortLab — Data Structures & Algorithms, Operating Systems, Machine Learning

    2026 — Present

  • Teaching Assistant — Scalable Data Science

    NPTEL

    2026 — Present

  • Teaching Assistant — Machine Learning

    IIT Gandhinagar

    2026 — Present

  • Teaching Assistant — Data Structures & Algorithms

    IIT Gandhinagar

    2026

  • Teaching Assistant — Computing

    IIT Gandhinagar

    2024 — 2025

Get in touch

Open to conversations on edge AI, efficient inference, and split computing — for research collaboration or otherwise.

yash.sahu@iitgn.ac.in