Currently leading the face anti-spoofing team where we leverage deep learning for fraud detection during KYC onboarding. Working on making AI models robust against level-1 and level-2 presentation attacks, deepfakes and Image Injection. Leading talent hiring for the AI team by harnessing the power of LLMs.

Built robust PAD (Presentation Attack detection) algorithms to identify 3D masks, latex or silicone masks, and other artefacts that are significantly harder to detect, leading to attainment of ISO 30107-1/30107-3 Level 2 compliance certification with 0% FAR (False Acceptance Rate) and 0% FRR (False Rejection Rate).Developed single-image-based deepfake detection models that are robust to complex deepfakes with an accuracy of 98% using a single image capture - anchored the data collection process, and literature survey, trained and experimented with several AI models and deployed the model into the production pipeline.Contributed to the complex challenge of image injection by developing an anomaly detector model that stops a large chunk of injections. Reduced the number of deepfake injections from 200 per week to 0 within a month by setting up a robust pipeline.Responsible for setting up and leading a DB-cleanup activity filled with fraudulent users - Developed an end-end pipeline using Docker consisting of custom AI models that could process 400k images per day.Mentored 3 interns to help and support them successfully achieve their goals. Worked closely with the engineering team to take our models to production. Also worked with product and business teams to understand client requirements and product ideation.

Worked end-end on the development of synthetic face identification module - explored large datasets like Face Synthesis, trained several SOTA deep learning models like ResNet, EfficientNet and VIT. Developed a lightweight eyewear detection model that was deployed to the production pipeline with an accuracy of 99.8% accuracy and an inference time of 5.8ms.

- Worked on shape analysis of complex brain surfaces by using Graph CNNs and asymmetric spectral filters.- Developed an end-end python pipeline for spectral embedding of brain surfaces. (Code: https://github.com/kharitz/aligned_spectral_embedding)- Mentor: Dr. Hervé Lombaert (Professor, Research Scientist - ETS Montreal)

1) Worked on Federated learning for medical data to enhance data privacy without compromising on the efficiency/accuracy.2) Created a federated learning simulation with 2 clients using flower platform where client-1 has liver and tumor data and client-2 has only liver scans.3) Mentor: Dr. Renuga Kanagavelu (Scientist II, Institute of High Performance Computing (IHPC))

- Worked on semantic segmentation of liver vessels using CT data to diagnose patients with a particular disease. Developed a preprocessing pipeline for contrast enhancement and resampling of CT images.- Experimented on U-Net models to accurately segment minute vessel structures.- Did extensive literature survey on existing techniques to perform the vessel segmentation.

Contributed to Mapping illegal dumpsites using AI along with 50 other collaborators across the world.Was part of Trash classification Task where we developed a dataset, annotated the data, and used state of the art segmentation techniques for better accuracy.