Krishnatheja Vanka

Help the discord community by solving queries related to Machine Learning and Deep Learning. Interact with fellow experts and share knowledge with each other while conveying the same to the members of the community.

Owned the end-to-end training lifecycle, including processing raw sports videos, generating and validating annotations using AWS Ground Truth, and fine-tuning and evaluating YOLO-v9, RT-DETR, and Grounding DINO to improve detection accuracy under diverse camera angles, occlusions, and lighting conditions. Applied and benchmarked multi-object tracking algorithms using BoxMot to maintain consistent player identities in highly occluded and crowded scenes, analyzing trade-offs between ID stability and computational cost. Designed and deployed scalable ML inference pipelines on AWS using ECS for service orchestration, Amazon ECR for containerized model artifacts, SageMaker for managed, asynchronous model serving, and MLflow for experiment tracking and model lifecycle management.

Conducted a comprehensive literature review, developed and trained a model from scratch, tested and deployed weather classification models, achieving >98% precision and recall, enhancing visual condition recognition for edge-based safety systems. Applied research-driven active learning methods (e.g., Influence Selection) to reduce data labeling effort by 40%, while maintaining >90% across all key metrics (accuracy, precision, recall, F1). Leveraged distillation and representation learning models such as DINOv2 and CLIP, integrated with ScaNN and One-vs-All retrieval, to automate dataset curation for weakly labeled and unlabeled data. Led end-to-end research and development of a driver fatigue detection module using a MobileNetV2 + GRU architecture; achieved 94%+ precision/recall for real-time yawn detection across diverse lighting and camera angles. Architected distributed training infrastructure across multi-node, multi-GPU clusters to scale training on 10B+ image datasets, enabling the development of proprietary vision models. Accelerated model development cycle by 15x by converting legacy Keras/TensorFlow pipelines into modular PyTorch Lightning framework, significantly improving reproducibility and maintainability. Routinely worked with AWS infrastructure (S3, EC2, SageMaker, EKS, ECR) to support scalable training, deployment, and monitoring of production models.

Developed a pytorch based framework with configurable training pipelines for object detection, segmentation, and vision transformers within Mareana's Manufacturing Intelligence Hub, reducing manual setup time by 70%. Researched and trained a transformer-based model to convert chemical equation images into corresponding formulas. Applied self-supervised learning techniques (DINO, Segment Anything) to reduce annotation cost by 40%, accelerating dataset preparation for downstream tasks. Optimized model inference latency by 3x by converting TorchScript models to TensorRT via ONNX, enabling real-time performance in production environments. Developed and maintained RESTful APIs using Flask and FastAPI for seamless integration of ML models into web-based applications and internal tools. Mentored junior ML engineers, guiding them through complete ML pipeline development including data acquisition, annotation (Labelbox), model prototyping, evaluation, and deployment best practices.

Developed image-processing prototypes that secured funding during Lightbox investor meetings, showcasing real-world applications of computer vision in agri-supply chains. Designed and deployed de-identification models to censor license plates and faces, enabling the rollout of AI-powered distribution centers while maintaining compliance with data privacy standards. Implemented YOLOv4-based object counting system to track vegetable quantities on conveyor belts, improving inventory accuracy by >85%. Built and managed end-to-end computer vision pipelines using Labelbox for annotation and MLFlow for model tracking and reproducibility. Applied quantization and pruning techniques to reduce deep learning model sizes by up to 60%, enabling low-latency inference on edge hardware.