● Open to Work

Ashpak Jabbar Shaikh

Deep Vision and Model Optimization Intern
Pune, India

About Me

CV Engineer with a 58% EPE reduction on a production robotics dense correspondence model (UFM/DINOv2), achieving a new SOTA of 1.40 EPE via DoRA + STRING RoPE, outperforming full encoder fine-tuning while reducing trainable parameters by over 90%. Open-source contributor to Google Keras Hub and experienced in developing production DICOM AI pipelines.

Skills & Technologies

Dense CorrespondenceOptical FlowDINOv2UFMRoPE VariantsSemantic SegmentationObject DetectionDepth EstimationMulti-View GeometryPinhole Camera ProjectionLoRADoRAPyTorchTensorFlowKerasOpenCVHugging FaceScikit-learnONNXTensorRTTritonDICOM Processing3D MRI/CT SegmentationOOD DetectionAWSDockerNVIDIA Container ToolkitWeights & BiasesMLflowDVCHydraTPU/Mirrored StrategyAsynchronous CheckpointingLangChainRetrieval-Augmented GenerationTransformersFAISSChromaDBNeo4jPythonSQLGitLinuxGoogle Cloud StorageMongoDBStreamlitTensorBoardAviskar State FinalistDeepLearning.AI ScholarMeta Certified Professional

Work Experience

Deep Vision and Model Optimization Intern

Perceptyne
Feb 2026NOW

Achieved a 58% EPE reduction through extensive ablation studies on UFM, discovering optimal feature extraction points and validating performance improvements. Implemented DoRA and STRING RoPE techniques to achieve a new state-of-the-art EPE of 1.40 while significantly reducing trainable parameters. Resolved critical MLOps bottlenecks by architecting efficient data pipelines.

DICOM AI/ML Engineer Intern

Unidoc Healthcare
Nov 2025Feb 2026

Engineered a safety monitoring layer for live DICOM streams and implemented algorithms for data drift detection to prevent model failures. Developed a validation pipeline for 3D medical segmentation models to ensure compliance with regulatory standards.

Projects

End-to-End 3D Medical Segmentation Framework for the Medical Segmentation Decathlon

Developed a config-driven MLOps framework supporting multiple 3D architectures, implementing custom training strategies to address class imbalance and achieving high performance on medical segmentation tasks.

TPUUNET++TransUNetSwinTransUNet

AI Study Assistant: Personalized LLM-Powered Tutor

Created a retrieval-augmented generation system that transforms handwritten and typed notes into an interactive study platform, leveraging advanced models for contextual Q&A and automated quiz generation.

LangChainHugging FaceOCRFAISS+1

Built a TPU-Distributed Semantic Segmentation Benchmarking Framework

Implemented various segmentation models from scratch on the Cityscapes dataset using a modular TPU-distributed training pipeline, developing custom loss functions and evaluation metrics.

FCNUNetUNet++

Education

BE

Artificial Intelligence & Machine Learning

Progressive Modern College of Engineering2022 - 2026
Ashpak Jabbar Shaikh - Deep Vision and Model | HiringAnt