Software developer(Contract) – Caltrans (California Department of Transportation), San Jose, CA

Project: Real-Real Time Traffic Analytics and Safety Monitoring System

Conducted applied research on computer vision and multimodal traffic datasets, designing and

fine-tuning deep learning models in Python + OpenCV to identify accident patterns and traffic anomalies.

• Designed structured evaluation frameworks to compare model performance across distributed inference pipelines, improving interpretability and reproducibility of results.

• Developed data ingestion and synchronization workflows using Kafka, Redis, and PostgreSQL,

enabling controlled experiments on large-scale video streams.

• Documented findings and system behaviors throughexperiment reports, error analysis summaries,

and annotation logs to support data-driven decision-making.

• Integratedvision-language embeddings (CLIP)with object tracking methods (ByteTrack) for occlusionhandling research, producing measurable accuracy gains in context-aware tracking.

• Collaborated with faculty researchers to prepare briefs summarizing key metrics, drift patterns, and

improvement strategies for model retraining cycles.

Conductedsatellite imagery analysis and model evaluation usingPyTorch, TensorFlow, andApache

Spark, supporting predictive studies on environmental and yield forecasting datasets (20TB+).

• Built research-ready data processing and orchestration pipelines (Airflow + Spark) to ensure consistent model retraining and reproducibility in geospatial experiments.

• Assisted in developing internal evaluation dashboards (FastAPI + Django) to visualize experiment

outcomes and metadata for ongoing research projects.

• Created annotated datasets, benchmark reports, and process documentation to support long-term

studies in geospatial model validation.

• Integrated Oracle + HDFS validation mechanisms for ensuring data quality and lineage across research experiments.

Project 1:Automatic Number Plate Recognition System

Conducted applied research on object detection (YOLOv3) and OCR-based recognition for automatic number plate detection and masking in real-time video streams.

• Evaluated the impact of image preprocessing and environmental variation on model performance,

publishing structured error analysis findings for internal use.

• Supported data annotation and curation efforts for large video datasets to improve model generalization across diverse weather and lighting conditions.

• Assisted in building analytical dashboards (Power BI + React) to visualize detection performance and

research metrics for stakeholders.

• Authored reproducible experiment logs and documentation to support further investigations in privacycompliant visual analytics.