Aug 2025 – Jan 2026 · 5 mos · California, United States

• + Architected low-latency video-to-3D point cloud reconstruction pipelines for autonomous FPV navigation using SLAM/SfM, optical-flow, FOE estimation, depth-from-motion, and TSDF fusion with CUDA-accelerated inference and real-time ROS2 integration.
• + Built a large-scale FPV dataset pipeline (web-scale crawl, metadata normalization, deduplication, artifact rejection) with automated quality filters via custom optical flow consistency, field-of-expansion stability, blur/compression scoring, and motion priors.
• + Trained a domain-specific vision foundation model for monocular ego-translation and pose estimation using self-supervised objectives (photometric reprojection, geometry consistency, cycle losses) and large-batch distributed training with ONNX/TensorRT deployment.
• + Integrated perception-to-control by fusing ego-motion, 3D occupancy, and risk fields into navigation stacks (local planning, collision avoidance, failsafes), achieving real-time throughput on edge GPUs with telemetry and health monitoring hooks

Jul 2024 – Jul 2025 · 1 yr · Downtown Toronto, Canada

• + Developed multimodal generative & understanding models optimized for wafer-scale chips (CS3), working on scaling up experiments
• + Created a transformer-based backbone for predicting interleaved visual and textual tokens, facilitating improved performance in visual question answering and image generation tasks, enhancing training/inference efficiency
• + Supporting advanced diffusion-based video generation models on wafer-scale chips, addressing memory bottlenecks and reimplementing existing architectures to optimize for weight-streaming, thereby enhancing model deployment on specialized hardware

Aug 2022 – Jun 2024 · 1 yr 10 mos · Bangalore Urban, Karnataka, India

• + Developed novel uncertainty-aware deep-learning regression models to work with imbalanced time-series data like ECG, that deduce disease states.
• + Core developer for the first version of a chat platform (among flagship products by Nference), integrating LLMs with in-house ML models, creating a platform that streamlines patient and medical exploration for physicians.
• + Developing novel method to produce 3D cardiac models from intra-cardiac-echo, harnessing GenAI.

May 2022 – Aug 2022 · 3 mos · Germany · On-site

• + Made monocular depth estimation efficient to allow real-time inference on autonomous robots
• + Designed and implemented a new uncertainty-aware test-time-training depth estimation scheme
• for robust adaptation to novel environments in an efficient fashion improving latency by 60%

Jul 2021 – May 2022 · 10 mos · Greater Tubingen Area

• Anomaly detection using computer vision techniques to optimize the production lines in manufacturing companies by automating inspection processes. Design, implementation, and deployment of the optimized model on low-resource devices with low latency.

Sep 2020 – Dec 2020 · 3 mos · Bangalore Urban, Karnataka, India

• Worked on low-cost smartphone-based portable retinoscopy device to diagnose several eye-related pathologies in a fast, convenient, and cheap manner.
• Designed and implemented an android app to capture the demographic, imaging, and IMU sensor data using smartphones and retinoscopes. Derived the mathematical formulation for the retinoscopy using data from smartphones to obtain the refractive error of the human eye. Implemented image-processing, sensor-fusion, and machine learning methods and deployed them on low-resource smartphones.

Oct 2019 – Sep 2020 · 11 mos

• Developed and deployed Bayesian key point detection models for radiographs (MRI) of the human spine and a feature to prioritize the MRI scans to be reviewed by expert radiologists based on uncertainty in the prediction.

Apr 2019 – Oct 2019 · 6 mos

• Developed machine-learning algorithms to filter and assign weights to thousands of proprietary return forecasts of stocks across various markets in USA, Canada, Europe, and Japan in Trexquant’s database.

Nov 2018 – Jan 2019 · 2 mos

• Worked on deep learning methods for spinal MRI scans

May 2018 – Jul 2018 · 2 mos · Tokyo, Japan

• Worked on deep learning methods for NLP for tasks such as NER @ HRI Japan

Nov 2017 – Jan 2018 · 2 mos · Mumbai, Maharashtra, India

• Developed real time face recognition system using multi task convolutional networks

May 2017 – Jul 2017 · 2 mos · Singapore

• Worked on side channel attacks for block ciphers

Nov 2016 – Jan 2017 · 2 mos · Banglore

• Implemented unsupervised clustering algorithms in JULIA