Projects

Agentic product search platform for architects. Microservices architecture with 8x crawling throughput, multi-agent workflows, VLM fine-tuning, Neo4j knowledge graphs, and AWS EKS autoscaling.

Fashion search API with finetuned CLIP embeddings, semantic similarity search, and real-time segmentation. Features sub-10ms vector queries, MCP integration, and enterprise-scale performance.

AI-powered cultural cooking platform with sequential Langchain workflows, RAG architecture, and DSPy-optimized prompts. FastAPI backend, MongoDB storage, validated with 44-user testing.

Autonomous exploration robot with SLAM, frontier navigation, and advanced path planning. Features K-means clustering, A* planning, and MPPI control. TurtleBot3 validated.

AI-powered calorie tracking with custom-trained object detection models for automatic food recognition. Features real-time classification, scale integration, and robust performance across diverse food presentations.

Ultra-low latency object detection and tracking using classical computer vision techniques. Features region-based extraction, distance-based matching, and millisecond-level inference for real-time performance.

Multi-modal image retrieval system combining color, texture, spatial, and deep learning features. Implements feature weighting, fusion strategies, and similarity metrics for semantic search.

Urban LiDAR mapping using modified LEGO-LOAM SLAM on NUANCE vehicle. Processed Velodyne VLP-16 data across Boston streets for detailed 3D mapping.

Sensor fusion study of IMU and GPS data from NUANCE vehicle on Northeastern campus. Developed algorithms for accurate pose and velocity estimation.

MPPT research comparing Perturb & Observe vs ANN methods for solar PV systems. Improved tracking efficiency and stability, IEEE published with boost converter.

Road surface segmentation research improving cross-dataset generalization for autonomous driving using novel training methodology with domain-invariant features.

Traffic sign detection with custom region proposal (5x fewer regions than selective search). IEEE published, optimized for real-time driver assistance systems.