Built a local Retrieval-Augmented Generation system

Role: SensorSpec AI Assistant (RAG System)

Customer: GenAI & LLM

Tasks:

• Built a local Retrieval-Augmented Generation system using LangChain and ChromaDB for querying large PDF document collections.
• Implemented structured document ingestion (PDF ? Markdown) to improve chunking and retrieval accuracy.
• Integrated LLaMA-based LLM via Groq API.
• Improved answer relevance and reduced hallucinations through a chunking strategy and two-step queries.

Engineered a system that analyzes acceleration

Role: Human Activity Recognition

Customer: Ubiquitous System development

Tasks:

• Publication: Accepted by iWOAR 2024 conference and published in Springer Forum.
• Engineered a system that analyzes acceleration & Pressure data to predict user activity (Lifts vs. Stairs).
• Implemented Random Forest classification algorithms to validate the use of pressure data in HAR tasks.

Designed a novel approach

Role: Multi-Modal Hand Pose Estimation (Master's Thesis)

Customer: Computer Vision

Tasks:

• Designed a novel approach using DeepConvLSTM architecture with a fusion layer for multi-modal development.
• Combined IMU sensor data from gloves with video input for multi-class gesture classification and pose reconstruction.

Skills:

Deep Learning, Mediapipe, OpenCV, CNN, LSTM