Highlighted Projects

Foundation Deep Learning Model for CryoET Python · PyTorch
Dec. 2025 – Present
  • Working on a foundation DL model for object detection and segmentation in CryoET 3D tomograms.
  • Gathered ~1,600 3D tomograms and generated 350,000+ 3D patches for pretraining using the 3DINO self-supervised learning approach.
  • Finetuning the model on different downstream tasks and benchmarking against baseline models (in progress).
Agentic AI System for Digital Pathology Python · PyTorch · LangChain · VLM
Nov. 2025 – Present
  • Engineered a multi-modal agentic framework orchestrating custom tools for automated WSI analysis, including Visual Question Answering (VQA), report generation, and zero-shot classification and ROI segmentation.
  • Integrated state-of-the-art pathology foundation models (SlideChat, TITAN, CONCH, MUSK) to enable context-aware reasoning over gigapixel-scale images.
  • Implemented a human-in-the-loop interface with a hierarchical k-means algorithm to iteratively refine model predictions, ensuring interpretability and clinical alignment.
MAPL3 — Mapping Axonal Projection in Light Sheet Microscopy in 3D Python · Bash
Jan. 2023 – May 2025
  • Developed a 3D computational pipeline for identifying individual axonal fibers and their brain-wide circuitry at single-fiber resolution.
  • Pioneered a novel DL architecture combining convolutional neural networks and vision transformers.
  • Deployed self-supervised approaches to improve model generalizability and reduce dependency on labeled data.
  • Designed and implemented parallelized image processing pipelines for pre- and post-processing [code].
ACE — AI-based Cartography of Ensembles Python · Fiji · Bash · Docker · Git
Jan. 2020 – Jan. 2025
  • Designed and implemented an end-to-end 3D pipeline for mapping local cell activity in tera-voxel scale (~1TB/sample) light sheet microscopy datasets.
  • Engineered an ensemble of convolutional and vision transformer-based models for soma segmentation.
  • Integrated Monte Carlo dropout for probabilistic model uncertainty estimation.
  • Optimized cluster-wise permutation statistical analysis for high-dimensional whole-brain microscopy data.
Foundation Deep Learning Model for Light Microscopy Python · Bash · Git
Jan. 2023 – Dec. 2025
  • Working on a foundation DL model for object segmentation across microscopy datasets and modalities.
  • Gathered ~3 million 3D image patches from multiple centers and modalities for pretraining using the 3DINO self-supervised approach.
  • Finetuning for axon, soma, and vascular segmentation; benchmarking against state-of-the-art models.
EchoJEPA — Latent Predictive Foundation Model for Echocardiography Python · PyTorch · Self-Supervised Learning
2024 – Present
  • Contributed significantly to the development of EchoJEPA — a latent predictive (JEPA) foundation model for echocardiography pretrained on 18 million videos across 300K patients — through technical implementation, experimental design, model evaluation, and interpretation of research findings.
  • Collaborated closely with an interdisciplinary team on methodological development, validation strategies, and scientific discussions to ensure robust and clinically meaningful outcomes.
  • Contributed to manuscript preparation, figure generation, and scientific narrative refinement; model achieves ~20% improvement in LVEF estimation and 78.6% view classification accuracy using only 1% of labeled data [website].
Pancreas Multi-Task Learning with nnUNetV2 Python · PyTorch · Git
Apr. 2025 – Jun. 2025
  • Developed a DL pipeline for joint pancreas segmentation and subtype classification on abdominal CT using a modified nnUNetV2 backbone.
  • Integrated a custom classification head with multi-scale feature adapters and fusion modules processing encoder outputs alongside segmentation decoding.
  • Achieved average Dice of 0.90 (pancreas) and 0.60 (lesion), macro F1 of 0.81; reduced inference time by 24% via optimized prediction loop.
System for Recording and Analyzing Pulse Signals MATLAB
Sept. 2015 – Jun. 2016
  • Designed a system for continuous blood pressure measurement using PPG recorded from wrist and fingertip.
  • Developed an MLP neural network to estimate SBP/DBP with mean absolute errors of 4.94 mmHg and 4.03 mmHg.
  • Implemented feature selection algorithms including moving-backward and genetic optimization to identify the most significant signal features.