Anuj Kumar

Developed an AI-driven talking head avatar system for seamless lip-syncing, reducing video production time by 50% and enhancing virtual communication quality for over 10,000 users. Developed an Inpaint Anything model with 34 PSNR which uses text to remove an object in an image using Diffusion and Grounding DINO model. Collaborated with data team to test 200+ videos to check robustness of the model.

Managed data team to ensure 6000 data curation applying MLOPS and contributing to a robust end-to-end model pipeline incorporating HIPAA compliance for real-world deployment. Successfully managed website team to develop a website to let users visualize 3D models and perform end to end prediction. Successfully trained a Transformer model on a massive 15GB point cloud dataset, achieving 93.1% F1 score in 3D point cloud completion task. Leveraged the efficiency of Deepspeed to decrease training time 4 times and smoothly deployed the model on Google Cloud Platform.

Developed a personalized image-to-image translation model for a car catalogue using the Dreambooth Stable Diffusion model, resulting in studio-like images. Employed prompt engineering techniques like prompt matrix in InstructPix2Pix and ControlNet Stable Diffusion models to get 25000 studio like car images. Created a real-time object detection model on edge using YoloV7, trained on 80000 images achieving a 72.3% MAP for inspecting car damages.

Contributed to open source nnabla - a deep learning framework. Enhanced the efficiency of using Explainable AI by 90% by developing XAI API. Reduced data visualization of XAI application in CNN by 80% by modifying Saliency, SHAP and Integrated Gradients explanation methods. Improved the accuracy of Resnet, Resnext, WideResnet, and Densenet by 20% by integrating it with Attention Branch Networks.

Built a computer vision model that can predict vital signs with 89.22% F1 score using face detection and signal processing methods.

Trained 3DCNN and LSTM on 400+ videos on colorectal surgery and enhanced the accuracy of surgical phase recognition to 84.66%.