Hyderabad, India

Arnav Deshpande

I make large neural networks fit, and run fast, on small chips.

Machine Learning Engineer at Qualcomm AISW
M.Tech CSE, IIT Madras

See my work Get in touch
fake_quant(x, bits=3)
levels
8
vs. float32
10.7×
rmse
0

Illustrative: uniform quantization of a toy signal, the core trade-off in my day job.

About

Placeholder for a portrait of Arnav Deshpande

I am an ML engineer focused on compiling deep learning models to intermediate representations for SoCs and AI silicon. At Qualcomm I bring neural networks, including LLMs from the Llama 3.1 family, to Snapdragon NPUs, tuning them for accuracy, latency, memory and power.

My M.Tech research at IIT Madras pushes the same problem further down the stack: hardware-aware compiler optimizations, profile-guided specialization and JIT compilation for deep learning inference.

Outside of systems and compilers, I play the piano and harmonium, and produce audio and video content.

Research interests Edge AI and on-device inference, model quantization, deep learning, NLP, computer vision, bioinformatics

Experience

Qualcomm AISW Hyderabad

Dec 2024 to present

Machine Learning Engineer

  • Optimize and enable complex neural networks, including LLMs such as the Llama 3.1 family (3B to 8B) and GPT-3, for on-device inference on Snapdragon chipsets and NPUs, across CV, NLP, GenAI and multimodal models.
  • Apply the full range of quantization techniques, including post-training and quantization-aware approaches, alongside accuracy tuning and system-level optimization of latency, memory and power.
  • Work across the QNN SDK, QAIRT Converter and ONNX Execution Provider toolchains, and led priority delivery initiatives on the ONNX EP and QNN-QAIRT Converter workstreams.
  • Apply LLM-specific optimizations: KV-cache quantization (INT8/FP8), weight-only quantization (W4A16/INT4), attention-head optimizations and runtime memory footprint reduction.

Jun 2023 to Dec 2024

Machine Learning Engineer, Associate

  • Built model conversion pipelines and enabled early-stage neural networks on Qualcomm's AI Engine, across transformer and sequence-based architectures.
  • Enabled deployment of CV, NLP and LLM workloads through quantization, model conversion, debugging and customer enablement.

Nov 2022 to Jun 2023

Interim Engineering Intern

  • Wrote automation scripts for internal workflows and model conversion, working with TensorFlow, PyTorch and ONNX.

MINRO Lab, IIIT Bangalore Bengaluru

May 2022 to Jul 2022

Research Intern

  • Measured statistical similarity across VGG16 feature maps using ORB, SSIM, t-SNE and Euclidean distance, aiming to cut compute by filtering redundant layers. Code
  • Explored applications of federated learning in intelligent systems.

Entrepreneurship Cell, IIT Kharagpur Remote

Apr 2020 to May 2020

Machine Learning Intern

  • Built a drug-discovery pipeline applying SVMs and Bayesian algorithms, at the intersection of computational biology and machine learning.

Mar 2020 to Apr 2020

Cyber Security Analyst

  • Built a real-time lateral movement detection model using remote-sensing image encryption.

Projects

Placeholder thumbnail for the Adaptive Neural IR Compiler

M.Tech research, IIT Madras

Adaptive Neural IR Compiler (ANIC)

A unified-IR compiler framework for deep learning inference that combines profile-guided specialization, JIT compilation and multi-target backends for a "compile once, adapt anywhere" execution model.

I am investigating how constraint-based points-to and alias analysis can be integrated into IR lowering, to improve buffer management and reduce memory-aliasing overhead for LLM inference on heterogeneous hardware.

  • Compilers
  • JIT
  • PGO
  • Alias analysis
  • LLM inference
Placeholder thumbnail for MobileNet-v2 model compression

M.Tech coursework, CS6886W

MobileNet-v2 Model Compression

Quantization-based compression built from scratch, with no compression APIs, across 12 configurations from 2 to 8 bits on CIFAR-10.

compression
4×
size
8.53 → 2.13 MB
accuracy drop
0.03%
  • Quantization
  • PyTorch
View code
Placeholder thumbnail for the Chintan YouTube series

Audio and video

Chintan YouTube Series

Managed visual production, transcriptions and subtitle alignment for a 32-part radio series on YouTube.

  • Video production
  • Transcription
  • Subtitling
Placeholder thumbnail for the protein visualization project

High-performance computing

Protein Visualization (Bio-NMR)

A parallel pipeline that compares protein sequences by amino acid structure using OpenMP and MPI, cutting computation time by about 25.76% against a sequential baseline.

  • OpenMP
  • MPI
  • Bioinformatics
View code
Placeholder thumbnail for the music emotion classifier

Audio ML

Music Emotion Recognition

A two-level classifier: an ANN on 27 acoustic features separates energetic clips (about 92% accuracy), then a CNN on spectrograms assigns one of six emotions, reaching 85.3% overall. With Saloni Parekh, Y. V. Srinivasa Murthy and Shashidhar G. Koolagudi.

  • CNN
  • MFCC
  • Spectrograms
Placeholder thumbnail for biomedical named entity recognition

NLP

Biomedical Named Entity Recognition

A hybrid Bi-LSTM-CRF architecture with a multi-task learning framework, jointly optimizing named entity recognition across biomedical text corpora.

  • Bi-LSTM-CRF
  • Multi-task learning
View code
Placeholder thumbnail for the epigenetic clock project

Computational biology

Epigenetic Clock for Cell Ageing

A pipeline that estimates cellular age from DNA methylation index using linear regression, improving accuracy by 42% over baseline epigenetic clocks.

  • TensorFlow
  • NumPy
  • Pandas

Publications & honors

Publications

  1. Springer, Proc. ICCI 2022, pp. 393-406, Jul 2023

    A Modified LSB Steganography Algorithm to Store Images of Large Size

    GPU-parallel variants of naive LSB steganography. At 5-bit LSB, image similarity reached 96.04% against 61.90% for the sequential method. With Y. V. Srinivasa Murthy, Shashidhar G. Koolagudi, Saloni Parekh and J. Vaishnavi.

  2. IEEE ICTACS, Oct 2022

    Time Administration of Virtual File System Operations

    Proposed a VFS as an abstract overlay for heterogeneous file transfer, and benchmarked core file operations against a traditional file system.

  3. IEOM Conference, 2022

    Breast Cancer Detection Using Six ML Algorithms

    Benchmarked classifiers on the Wisconsin Breast Cancer dataset, reaching up to 96.50% diagnostic accuracy.

Patent

  1. Indian Patent Journal, 05/2022, App. No. 202241003742, granted

    Smart Stethoscope

    An IoT device that streams heartbeat data to the cloud for abnormal heartbeat detection and remote diagnostics, at about 32.5% lower cost than traditional stethoscope devices.

Honors

  1. Qualcomm, May 2026

    Impact Award

    For leading priority deliveries on the ONNX EP and QNN-QAIRT Converter initiatives.

  2. I-NEST Conference, Oct 2019

    Best Paper Presentation Award

    Ranked first in the Advanced Nanomaterials track for "Nanocomputing: The Future of Cancer Treatments through Advanced Drug Delivery".

Skills

Optimization & edge AI
Model quantization (PTQ/QAT), model conversion, inference optimization, accuracy tuning, LLM deployment
Frameworks & runtimes
PyTorch, TensorFlow, ONNX Runtime, TFLite, QNN SDK, scikit-learn, NumPy, Pandas
Languages
Python, C/C++, SQL, Bash, LaTeX
Research methods
High-performance computing (OpenMP, MPI), NLP, statistical modeling, bioinformatics
Tools
Git, Docker, automation scripting
Open source & mentorship
Contributed to the ONNX Runtime Execution Provider (ORT-EP) QNN repository. Mentored new hires, interns and junior lab members, and ran career and software engineering workshops at schools and colleges.

Education

Jul 2024 to Jun 2027

M.Tech, Computer Science and Engineering

Indian Institute of Technology Madras

Coursework: Modern Compilers, Advanced Algorithms, Program Analysis, High-Performance Computing.

Jul 2019 to Jun 2023

B.Tech, Computer Science and Engineering

Vellore Institute of Technology, CGPA 9.21/10

Coursework: DBMS, Operating Systems, Parallel Computing, Networking, Security, NLP, AI, Web Development.

Certifications

Digital Electronics and Logic Design (Udemy), Introduction to Programming with MATLAB (Coursera), Statistical Data Analytics (Business Research), Big Data Analytics (NASSCOM).

Contact

Working on on-device inference, quantization or ML compilers? I would like to hear about it.