Minh Tiến Trần Email & Phone Number

Ho Chi Minh City, Vietnam

BioStudio:- Managed all notebooks within the BioStudio platform, creating comprehensive and accessible tutorials for diverse bioinformatics methods including scRNA-seq, Bulk RNA-seq, and Spatial RNA-seq.- Isolated each notebook from its environment, resolving conflicts between R and Python packages, and providing corresponding data samples. Successfully handled multiple CUDA versions for notebooks utilizing GPU frameworks like TensorFlow, PyTorch, and AlphaSC, ensuring environment compatibility.- Created pipelines for running large-scale data analyses from Seurat (R) datasets with InferCNV and CellChat. These pipelines were implemented as tool scripts, allowing users to input datasets via the command line and automatically obtain all analysis results from the packages.Research in Cell-type Prediction:- Implemented memory lazy loading to handle massive-scale datasets exceeding RAM constraints during training. - Leveraged Variational Autoencoder models (scVI, scANVI), and attention-based models (scGPT, Geneformer) to enhance cell type prediction accuracy and integration performance. - While these models effectively mitigated batch effects, accurate prediction of rare cell types remains a challenge due to the inherent overdispersion and heterogeneous of scRNA-seq data, which is crucial for effective disease treatments.Research in Cell Instance Segmentation:- Researched and implemented a custom U-Net model from scratch, inspired by the Cellpose architecture for image instance segmentation. - Optimized model performance through architectural modifications, achieving a 3% improvement in mean Average Precision (mAP) compared to the original Cellpose. Accelerated inference by deploying the model using TensorRT, resulting in a 2x speedup. - While the model demonstrated improvements in specific scenarios, achieving over 90% accuracy remains crucial to fully meet the project’s requirements, as cell instance segmentation impacts all downstream analyses.

Ho Chi Minh City, Vietnam

Vehicle Counting:- Developed an algorithm to count the number of incoming/outgoing vehicles, including motorcycles, cars, etc., in multiple video streams and RTSP cameras using NVIDIA Deepstream.- Fine-tuned a YOLOv7 model for vehicle detection and improved the AI model's speed to achieve real-time performance by converting it to TensorRT's format. - Implemented vehicle tracking using the centroid tracker algorithm. The application detected and tracked vehicles in real-time with an accuracy of 98% accuracy.People Counting:- Developed an incoming/outgoing people counting and people alert zone in multiple video streams and RTSP cameras using Intel DLStreamer.- Fine-tuned a YOLOv7 model for people detection and converted the AI model to OpenVINO’s format to accelerate Intel’s CPU and GPU hardware while maintaining the same accuracy as the original model. - Implemented people tracking using the DeepSORT algorithm. The application detected and tracked people in real-time with an accuracy of 95% accuracy.Health Monitor OCR:- Led and supervised a team of interns in researching and implementing text detection, screen alignment by text orientation, text generation, and text recognition models using PaddleOCR. Worked closely with team members to ensure project goals and deadlines while providing guidance and support throughout the process.- Developed OCR model to capture information from health monitor devices with diverse schemes and designs based on YOLOv7 screen detection and PaddleOCR text recognition. - Proposed screen alignment by text direction classifier and templates to identify value labels by finding distance congruence between center points for each bounding box. Successfully deployed solution in company's production with 98% accuracy.