Debasish
Das.

Engineered a multi-agent stateful workflow using LangGraph to fully automate zero-shot enterprise RFQ processing. Implemented multi-vector search via Weaviate to achieve 87–90% precision on a 10k product catalog, incorporating an autonomous web-scraping layer to dynamically cross-reference real-time market valuations. Tracked and optimized agent trajectories using LangSmith to minimize token latency.

Architected a state-machine chatbot utilizing LangGraph to manage complex, non-linear user conversations. The system extracts structured user intents, routes queries to a Weaviate vector database for personalized gift matching, and executes a programmatic algorithmic layer to handle strict multi-variable budget constraints.

Designed an optimized data pipeline handling automated object detection across a 100-movie asset catalog. Built a semi-supervised labeling workflow to generate high-precision bounding boxes with minimal human overhead, benchmarking state-of-the-art YOLO and DETR models for real-time spatial analytics.

Prototype for cctv surveillance system with threat detection capabilities using YOLO, MediaPipe, and Transformer. Give real-time insights into safety metrics and potential threats.Send messages to authorities.

Hybrid ANN-CNN fusing tabular metadata and images. EfficientNetB0 backbone with dense layers achieving 88% accuracy and 0.2879 loss.

XGBoost model with feature engineering, 80% accuracy, deployed via MLOps practices on Amazon SageMaker with a Streamlit frontend.