Aug 2019 – Feb 2022 · 2 yrs 6 mos · Bengaluru Area, India

• 𝗠𝗲𝗱𝗶𝗰𝗮𝗹 𝗖𝗼𝗱𝗶𝗻𝗴: Predicting ICD (International Classification of Diseases) and CPT (Current Procedural Terminology) codes
• Led the development of AI based medical coding assistance systems to predict medical codes (procedural and diagnostic) from the patient encounter notes.
• Developed the NLP models to predict these medical codes with ~80% precision 65% recall for multi-label multi-class classification for a subset of labels covering ~80% of the codes encountered.
• Devised an interpretable classification method to highlight relevant portions of text from the patient encounter notes as evidence associated with predicted labels.
• Developed snowflake based SQL queries and Spark data processing pipelines for accumulating training data for ML model training for quicker model iterations.
• Implemented method to associate a level of confidence with the predicted medical code based on conformal mapping for prediction of medical codes.
• 𝗘𝗻𝘁𝗶𝘁𝘆 𝗘𝘅𝘁𝗿𝗮𝗰𝘁𝗶𝗼𝗻 𝗳𝗿𝗼𝗺 𝘀𝗰𝗮𝗻𝗻𝗲𝗱 𝗱𝗼𝗰𝘂𝗺𝗲𝗻𝘁𝘀
• Implemented a computer vision-based pipeline for identification and extraction of entities such as names, dates and cheque details from a scanned set of documents. This led to more robust detection of these entities as opposed to just using OCR software to extract.

Jul 2017 – Apr 2019 · 1 yr 9 mos · Bengaluru Area, India

• 𝐀𝐈 𝐩𝐚𝐭𝐡𝐨𝐥𝐨𝐠𝐲 𝐚𝐧𝐚𝐥𝐲𝐬𝐞𝐫𝐬
• Led the development of an AI based pathology analyzer to calculate the fertility metrics related to human semen samples.
• Solved tracking of cell movements (primarily based on Kalman Filtering) by associating multiple cells across a space of 60 microscopy video frames to generate cell trajectories. These trajectories were then used to calculate the overall mobility statistics as reported in a standard fertility report. Clinical studies against the gold standard (expert pathology labeled data)  resulted in a mobility correlation of ~0.81 and concentration correlation of 0.86.
• Replaced existing approach for cell localisation and identification from microscopy images with a single object detection model leading to faster analysis and simplified systems.
• Designed and implemented scalable evaluation workflows for various image processing and ML models involved, Additionally optimized machine learning training pipelines for efficient use of cloud resources.
• Developed slide focussing and scanning algorithms to reduce overall scan times from the in-house digital microscope from 10 minutes to 6 minutes.

Aug 2013 – Jul 2017 · 3 yrs 11 mos · Bengaluru Area, India

• 𝗥𝗲𝘁𝗮𝗶𝗹 𝗔𝗻𝗮𝗹𝘆𝘁𝗶𝗰𝘀
• Developed an ML-based object detection model to identify and track people from an overhead camera to measure a retail store’s footfall. The solution was deployed in nearly 100 stores across India during the first rollout. (including stores such as Sunglass Hut, Cafe Coffee day)
• Devised the optical flow algorithm to approximate depth from an uncalibrated stereo camera, thus saving on high calibration costs. This eliminated the need for a high-cost per-device calibration process.
• 𝗜𝗺𝗮𝗴𝗶𝗻𝗴 𝗦𝗼𝗹𝘂𝘁𝗶𝗼𝗻𝘀
• Led the development of a unique solution to stitch a panorama with HDR images. This solution enabled stitching multiple exposure bracketed shots (as opposed to general 8-bit stitching) into a panorama image.
• Adapted standard image-processing components to stitch and pyramid blend high dynamic range images (12-bit images), improving blending quality and more seamless stitching of two images.
• Reduced mobile device memory footprint brings it down from 130 MB to just 30 MB.
• Ported image-warping algorithms to run on GPU(s) on mobile devices using Renderscript for Android and Metal for iOS. This led to faster execution and a lesser memory footprint.
• Developed cross-platform HDR panorama stitching SDK in C++ for Android and iOS Devices.
• Developed a fully-fledged Panorama stitching demo application on Android to showcase the HDR panorama solution to the potential client.
• 𝗠𝗼𝗯𝗶𝗹𝗲 𝗔𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀
• Developed a closed caption decoder that could extract caption tracks embedded in video packets during the parsing stage in C++.
• Created a rendering component for Android that overlapped caption text during playback.
• Developing components to display network and video packet performance statistics during DASH (Dynamic Adaptive Streaming over HTTP) playback,
• Utilized OpenGL to develop video editing filters for Android media player application to demonstrate H.265 (the world’s first software video decoder for HEVC)