My research empowers machine learning models used in Billions! of devices (smartphones/laptop CPUs: I am interested in applying machine learning to real-world practical applications.

[LinkedIn]  [Scholar]  [My CV]
Email (sandha.iitr (at)

Research Themes: Develop state-of-the-art artificial intelligence systems by advancing fundamental research in multimodal sensing, novel machine learning algorithms, scalable deep learning, and lightweight machine learning for edge devices. My focus is to enable critical real-world applications in the areas of water monitoring, semiconductor designing, precision agriculture, and ocean monitoring.

What's Exciting :) 

Work Experience

 Fundamental ML Skills

 Leadership Roles

Professional Services


Selected Publications

Active Research Projects

More Details

1-Scalable and Robust Machine Learning

We introduce Mango, a novel state-of-the-art ML Library for parallel hyperparameter tuning. 

Mango is used in production to design ARM CPUs for IoT sensors to supercomputers, and from smartphones and laptops to autonomous vehicles.
Paper-1 (CogMI-21): PDF Slides  Paper-2 (ICASSP-20): PDF  Slides

2-Deep Reinforcement Learning based Active Tracking

We introduce a novel simulator developed over Unreal Engine controlling objects in photo-realistic virtual worlds. The simulator enables the training of deep-RL policies for tracking, outperforming traditional object detectors and Kalman baselines. 

Sim2Real: We show the transfer of models trained in simulation to the real world.  Paper: Accepted in IoTDI-23.

3-Robust Delay-aware Deep Reinforcement Learning

How do deployment variations impact state-of-the-art Deep-RL-based controllers?
We introduce delay-aware deep reinforcement learning and show its transfer to real-world robotic applications. We use physics-based simulations (Gazebo & PyBullet) and real robots (1/18th scale autonomous vehicles) for experiments.  We show the transfer of models trained in simulation to the real world.  

Paper (CoRL-20): Code  PDF   Demo Video  Slides-Video  Slides

4-Deep Convolutional Bidirectional LSTM for Complex Activity Recognition with Missing Data, Springer (2021).

How does ML behave with runtime uncertainties? Code  PDF  Video

(3rd in the Cooking Activity Recognition Challenge out of 78 teams.)

5-Time Awareness in Deep Learning-Based Multimodal Fusion Across Smartphone Platforms, IoTDI -2020

Tell me how accurate is my smartphone's time? Does it impact my ML applications?

Sandeep Singh Sandha*, Joseph Noor*, Fatima Anwar, Mani Srivastava. Code  PDF

GoodClock (Code) library is used by the David Geffen School of Medicine to collect data with accurate timestamps from patients.

6-Exploiting Smartphone Peripherals for Precise Time Synchronization, ISPCS-2019

What is the best possible timing accuracy on modern smartphones? oh! yes, it is in microseconds.

Sandeep Singh Sandha*, Joseph Noor*, Fatima Anwar, Mani Srivastava. Code  PDF

7-Radhar: Human Activity Recognition from Point Clouds Generated through a Millimeter-wave Radar. Mmnets, 2019

We present end-to-end deep learning for sparse point clouds from Radar and introduce a new open-source dataset. 

code slides pdf

8-In-database Distributed Machine Learning, VLDB-2019 

Sandeep Singh Sandha, Wellington Cabrera, Mohammed Al-Kateb, Sanjay Nair, Mani Srivastava.  PDF 

US Patent

9-Heliot: Hybrid emulation of learning enabled IoT systems. 

Sandeep Singh Sandha, Mani Srivastava.  Code

Best Demo award at IoTDI 2019

10-Cross modal training for activity recognition. EdgeSys-2018 

Tianwei Xing*, Sandeep Singh Sandha*, Bharathan Balaji, Supriyo Chakraborty, Mani Srivastava.  Code PDF

11-MetroInsight: Data-hub architecture for smart cities

Code Paper Slides & Demo

MetroInsight was developed in collaboration with Microsoft Azure and the data ingestion service is hosted on the cloud.

Data-driven analytics for water: BlueWater & WaterWatch (IBM Research)

Project page PlayStore Github API Paper Paper

Processing health data in real-time using Apache Kafka and Spark (UCLA)

Project Page  and Git Code  Project Report 

Distributing TensorFlow graph computation (UCLA)

Project Page Project Presentation  GitHub 

Real-time query interface for large datasets using Apache Spark (UCLA)

 Slides , Project Report , GitHub 

We used mathematical modelling to numerically solve the partial differential equations. The system is a industrial scale large mesh solver that uses MPI environment to handle millions of meshes in 3D. (IBM Research)  

Publications:  Paper-1 (2017), Paper-2 (2016),  Paper-3 (2016), Paper-4 (2016), Paper-5 (2015). 

Smartphone based remote health monitoring system (IIT Roorkee)

This was my bachelor thesis project to develop smartphone based healthcare.
IEEE Healthcom 2014. 

Context-aware panoramic map generation using opportunistic crowdsourcing (IIT Roorkee)

CP360 generates a fully-tagged 360 degree panoramic map of the surroundings of a querying user using crowd-sourced images, audio trails, object tags, and raw location data collected by smartphones in an opportunistic manner.  IEEE TPDS 2014

IBM national technical challenge, 2013 (IIT Roorkee)

This was project done as part of IBM National Technical Challenge (NTC 2013) to predict crime prone areas and suspects. Our work was accepted among the top five teams in finals. 
Slides (Slides are released as they were!)