Apr 2023 – Sep 2023 · 5 mos · Bengaluru, Karnataka, India

Apr 2021 – Apr 2023 · 2 yrs · Bengaluru, Karnataka, India

• ◦ Worked for R&D of a leading AI Chip Developer Client.
• ◦ Actively participated in development of a tool which optimizes the Tensorflow models’ TensorArray operations and Loop based operations which often are not supported on the AI accelerators.
• ◦ Engaged in core designing, development and testing of a tool which takes any ONNX/TF model and applies various types of modifications and optimizations such that it can run efficiently on the AI accelerator chip. This includes analysing the model and identifying and modifying the graph based on the node patterns.
• ◦ Responsible for understanding and exporting of various Classification, Detection and NLP models for their model zoo. Developed necessary inference scripts along with post processing routines.
• ◦ Helped in debugging accuracy issues by analyzing the architecture of the models.
• ◦ Gave presentations to the internal team about intricate details of the architectures.
• ◦ Mentored teammates by having regular knowledge sharing sessions.
• ◦ Enabled new joiners for their upcoming assignments by providing them problem statements to work on and guided them.
• ◦ Assisted the hiring team by taking interviews during this period

Jan 2019 – Apr 2021 · 2 yrs 3 mos · Vadodara, Gujarat, India · On-site

• Researched and developed deep learning solutions integrated into core products like CCTV cameras and attendance devices, focusing on computer vision for edge deployment.
• Developed a proprietary License Plate Recognition model from scratch, achieving 86% end-to-end accuracy in production, a result competitive with leading solutions, and state-of-the-art (SOTA) accuracy for Indian conditions. Innovated a novel approach to solve the sequence prediction problem without RNNs/LSTMs, significantly reducing latency and model size for edge deployment.
• Engineered an end-to-end Face Recognition stack, designing and training the entire pipeline (detection, keypoint, alignment, recognition). Achieved a 99% recall rate at a <0.1% false accept rate for recognition and 99% precision/97% recall for liveness detection on in-house benchmarks. Models were trained on a proprietary dataset robust enough for challenging real-world deployment conditions.
• Designed a custom, edge-optimized object detection model based on YOLO architecture to detect key objects (e.g., persons, vehicles) in surveillance footage with challenging environmental conditions.
• Deployed models to production by converting and optimizing them for various SoC chips (e.g., using TensorRT, OpenVINO). Leveraged C++ APIs to build a production-grade SDK, ensuring optimal inference speed and seamless integration into commercial security products.

Jul 2016 – Sep 2018 · 2 yrs 2 mos · Surat, Gujarat, India · On-site

• Led cross-functional teams in the end-to-end execution of maintenance projects for a large-scale distributed network (380km), ensuring on-time and on-budget delivery.
• Pioneered a custom Android application to digitize and streamline the tracking of 650+ asset locations, enhancing operational efficiency.

Mar 2014 – May 2015 · 1 yr 2 mos

Sep 2012 – Mar 2014 · 1 yr 6 mos