Feb 2023 – Mar 2025 · 2 yrs 1 mo · Remote

• 1. AI chip company: Developing real-time low-power Computer Vision algorithms for augmented reality on smart glasses.
• 2. Content moderation company: Deploying Large Language Model (LLM) solutions to moderate multimodal data online.
• 3. Togal.AI : Building Computer Vision solutions for detecting, measuring and comparing project features on architectural plans and drawings

Oct 2021 – Present · 4 yrs 8 mos · Oxford, England, United Kingdom · On-site

• Research focus: 3D/4D Reconstruction and Generation, Vision-Language (Multimodal) Foundation models, 3D+LLMs
• Advisors: Andrew Zisserman, Andrea Vedaldi, Joao Henriques, Iro Laina.
• Publications at CVPR, NeurIPS, ECCV, ACCV, ICLR and 3DV.

Nov 2020 – Jul 2021 · 8 mos

Jun 2019 – Oct 2020 · 1 yr 4 mos

• Efficient Deep Learning for Computer Vision -- algorithm development and system design
• Spearheaded the ultra-low resource always-on vision project from model design, quantization to final hardware mapping
• Filed 12 inventions in 2020-21 of which 6 ideas have been filed for patent protection. Notable works on 3D hand-pose estimation, low-bit quantization, structured and unstructured pruning.
• Led Qualcomm’s team in the MicroNet Challenge at NeurIPS 2019, and achieved 3rd position in ImageNet track [https://github.com/yashbhalgat/QualcommAI-MicroNet-submission-MixNet]
• Managed/mentored interns - Jangho Kim and John Yang (PhD @ SNU) with contributions to the AR/VR project

Jan 2019 – May 2019 · 4 mos

• Researched and developed production pipelines for real-time vehicle tracking for querying on large-scale video databases

Jun 2018 – Aug 2018 · 2 mos · Almaden, San Jose

• Worked at IBM Almaden Research Lab with the Watson Languages group on task-agnostic classification in the presence of label noise
• Built ensemble-based frameworks for combining weakly-labeled (or mislabeled) and high-quality samples for the training of a sentiment model.
• Work accepted to KONVENS 2019

May 2017 – Jul 2017 · 2 mos · Paris

• Used Scattering Wavelet Networks for the classification an segmentation of seismic structural "monads". Work done has been accepted as a paper at ICASSP 2018. You can read about it here: http://www.laurent-duval.eu/opus-cats-eyes-seismic-data-classification-scattering-networks.html

May 2016 – Jul 2016 · 2 mos · Bangalore

• Used CorrNets, an autoencoder-based architecture, to learn the joint representation for images
• and captions. We were able to obtain state-of-art results for large fashion catalogues search without manual tagging.

Nov 2015 – Dec 2015 · 1 mo · Pune, Maharashtra, India

• With specific recognition to stamp detection and segmentation, we proposed a shape-based ranking
• algorithm to learn the 1st layer of a CNN. Detection accuracy 94% and segmentation IoU 74.81%. Work accepted as a short Paper at the DAS 2016 conference.

May 2015 – Jul 2015 · 2 mos · Mumbai, Maharashtra, India · On-site

• Software module development using CAD modelling engine
• Developed a range of ‘constraint-modules’ for automated construction of furniture parts using the FreeCAD engine.

Nov 2014 – Dec 2014 · 1 mo

• Designing an Indoor Navigation System using Pedometry and Particle filters.
• My work (Pedometry) involved on IMU sensors:
• 1. Using various algorithms like PCA, Triad algo to determine the heading of motion,
• 2. Use FFT, CWT etc. to determine the step frequency and model the step length of the user.

Aug 2014 – May 2015 · 9 mos

• Development of a Mars rover with Mars Society of India for the University Rover Challenge organized by Mars Society, Utah, United States.
• ● Worked on Sensor calibration and testing for navigation
• ● Also a part of the Image Processing subsystem for vision-guided navigation of the rover.