Leading the multidisciplinary Artifacts, Registries and Automations team

Helping teams build Generative AI products. Ad hoc consulting, management coaching, training and product development.

Led the NLP team within the Ingest & Extract Group at BenchSci. I drove the adoption of advanced generative AI technologies like GPT-3.5, GPT-4, and Llamav2/Mistral models into the ASCEND platform, transforming our natural language processing capabilities and conversational UX to elevate business applications.

Leading the development and delivery of advanced AI systems for the Canadian Armed Forces (CAF). Product portfolio include Recce (real-time video analysis platform with AI), XT-SHIMS (platform for NLP-based analysis and visualization of worldwide events in near real time) and WISRD (wildfire intelligence, reconnaisance and surveillance ISR platform).

Setup and hire a new multidisciplinary product team for Xtract Recce. Led the Xtract Recce product team, building Canada’s first advanced Full Motion Video analysis platform using AI.

Led multiple internal initiatives for ML workflow optimization and MLOps implementation. Built the technical foundation for the productization of Xtract Recce, one of the first Canadian Video Analytics platform with real time AI inference geared for the Armed Forces applications.

- Focus on low-level deep learning and generative networks research and experimentation to improve dermatological/clinical image classification.- Improve and optimize existing MLOps infrastructure reducing costs by 40% and response times by 50%.

- Developed and deployed a clinically relevant deep learning system for advanced pathology testing directly from MRI scans (100% non-invasive)- Collaborated with clinicians, neurologists, computer scientists and patients to understand requirements and propose implementation plan.

- Leading a team of developers in their software development efforts towards a revolutionary new electronic health record (EHR) system which leverages advanced natural language processing and document parsing technologies.Responsibilities include:- Working closely with the CEO Dr. Roberta Lee in identifying current problems in existing EHR systems, and brainstorming solutions for the same.- Leading a group of consultants in their work in front-end development, database creation and API design.- Implementing critical NLP/Document Parsing APIs on the backend using the Python/Flask framework that implements business logic, using a combination of public APIs and in-house tweaks.- Inventor of the internal PlutoHI DAG System, which greatly improves upon publicly available OCR APIs (Google Vision API) and provides more accurate document text detection. Invention is patent pending.Technologies used:- Python, Flask, Google Cloud Platform, Google Vision API, NetworkX, nginx.

- Implemented a number of state-of-art segmentation algorithms for segmenting tumor structures in brain MRI scans. -Wrote a complete modular framework from scratch that supports plug-and-play model definitions, image augmentation, multi-gpu training support and many other features. Link to open source framework: https://github.com/trane293/brats2017-proj/- Collaborated with Siemens Healthnieers for a project, where I contributed in the inception of idea, and implemented the proposed idea independently in Python/Tensorflow. Link to manuscript: https://arxiv.org/abs/1804.05181- Collaborated with the Single Molecule Localization Microscopy (nanometer scale microscopy) group at MIAL in a project that beats the state-of-art in determining proper region of interests of excitation compared to noise. Manuscript under preparation.- Interacted with collaborators at Vancouver General Hospital to understand, document, and propose CS-based solutions to the problems they face in their clinical workflow.Personal project:- Undertook a personal project independently for dealing with the problem of missing MR pulse sequences (missing input) when state-of-art segmentation models are deployed in clinical setting.- Developed a novel, state-of-art, multi-modal generative adversarial network (GAN) that synthesizes multiple missing MR pulse sequences by utilizing information from any number of available sequences. Link to manuscript: https://arxiv.org/abs/1904.12200 which is currently under review at a highly reputed journal in medical imaging.- Researched and presented state-of-art in brain tumor segmentation, generative adversarial networks, and image reconstruction throughout my tenure at MIAL.- Invited to present research at Vancouver Imaging Network at Centre for Brain Health, UBC, and for MRI Researchers’ Retreat at UBC.Technologies used:- Python, SimpleITK, Scikit-Learn, Matplotlib, t-SNE, Flask, PyTorch, Keras, Tensorflow, Flask, NiBabel, NiftyNet.

- Working together with Stanford/Caltech alumni towards developing a computer aided detection/diagnosis system for mammography (breast X-Rays) using state-of-art machine learning and deep learning models.- Developed the data processing pipeline from scratch in Python, that handles data coming from a clinic’s PACS system and converts it to a format that allows effective analysis.- Researched, designed, implemented and tested various classical machine learning (artificial neural networks, support vector machine, random forest classifier, decision tree) as well as state-of-art deep learning methods (convolutional neural networks) for detection of abnormalities in mammography images.- Implemented various convolutional neural networks architectures ranging from autoencoders, siamese networks and FCNNs using Python/Keras/Tensorflow stack.- Implemented classic machine learning based pipeline that extracts shape/textural/contrast features from image ROIs, performs feature selection, trains various ML classifiers and validates the classifiers using a held-out test set.- Developed pipelines to visualize network predictions, activation maps, feature spread (using t-SNE).- Interviewed multiple candidates for the role of Machine Learning Engineer at the company, where the candidates ranged from advanced undergraduates to graduate students from top universities in USA.- Winners of the TiEcon 2017 - TiE 50Technologies used:-Python, PyDicom, SimpleITK, Scikit-Learn, Scikit-Image, XGBoost, Matplotlib, t-SNE, Keras, Tensorflow.

- Undertook bachelor's dissertation research under Prof. Sushmita Mitra at Machine Intelligence Unit, ISI Kolkata. - Worked on the application of deep learning (convolutional neural networks) towards detection of tumors in brain MRI scan. - Proposed a combination of two CNNs, C-CNN and D-CNN for the classification of abnormal slices and detection of tumors in slices respectively. - Manuscript available on arXiv: https://arxiv.org/abs/1806.07589

Member of the Driverless Car Project team at the Cube26 Automotive Research department. The five member team is one of the 12 finalists for the $1 Million prize at the Mahindra Spark the Rise: Driverless Car Challenge.Initiated and lead the Computer Vision and Machine Learning subgroup for Cube26 AutomotiveResearch.Initiated research activities in the real time object detection and recognition domain. The CV/MLgroup designed a traffic lights detection and recognition module capable of working in real time.Ported earlier developed modules to GPU for fast and near real time performance.

Developed Traffic Sign Detection and Pedestrian Detection modules as a part of team working ondeveloping an Autonomous Car for Mahindra Spark The Rise Challenge. Researched the applicability of various shape and margin based features for sign detection. Also experimented with different classifiers like artificial neural network, support vector machine and k-nearestneighbors to find optimal feature + classifier combination with best performance.Was also responsible for porting and optimizing all the modules developed by team to GPU using NVIDIA CUDA C, as a result successfully improved performance of each module by at least 200%.

Implemented a Face and Eye detection module for the project Avyam Tryon in C++ using OpenCV. Also implemented a face highlight detection algorithm. Prepared case study report of various virtual try on applications on the web.

Developed a novel Computer Aided Diagnosis (CADx) system for mammography under theguidance of Dr. Jayasree Chakraborty and Dr. Abhishek Midya. The system was designed to automatically classify benign and malignant masses in mammograms and assist radiologists in the crucial decision making process. A novel and robust approach was proposed that makes use of invariant Zernike moments as features in the feature extraction stage of CADx system for the classification of masses. The system achieved a classification accuracy of 96.7% which was the highest ever reported using the MIAS database.

Assessed the performance of various spatial domain filters in denoising a grayscale imageunder the guidance of Dr. Jagroop Singh. Performed detailed survey and analysis of the available filters and carried out experiments to determine the best filter for a number of known noise models. Also tested the robustness of each filter on images corupted with more than one noise type and with noise of unknown probability distribution. Presented the work at IEEE 6th International Congress on Image & Signal Processing (CISP 13’) held in Hangzhou, China 16-18 December 2013 sponsored by IEEE EMBS Society.