My Projects

Project Overview:

A hybrid multi-agent research architecture that uses a local Qwen 3 0.6B model, served via Ollama, for planning, search orchestration, and research synthesis, while reserving Gemini exclusively for final report generation. The design tests whether different agent tasks can rely on different model capabilities instead of running a frontier model throughout the entire workflow.

Key Features:

• Multi-stage LangGraph workflow covering task decomposition, tool/search orchestration, research synthesis, and final answer generation
• Local Qwen 3 0.6B model handles planning, search orchestration, and synthesis
• Gemini used only for the final report generation stage
• Structured-output handling with JSON parsing and fallback mechanisms to improve reliability of smaller local models
• Explicit evaluation of the trade-off between local and frontier models across agent tasks

Results:

• Reduced API requests from 70 to 18 (~74% reduction)
• Reduced input tokens from ~150K to ~29.9K (~80% reduction)
• Increased observed success rate from 50% to 83.33%

Technologies Used: Python, LangGraph, Ollama, Qwen 3 0.6B, Gemini, Web Search

Project Overview:

An agentic RAG architecture in which a reasoning agent dynamically coordinates multiple information sources rather than relying on a single retrieval step. Evidence gathered from each source is fused together before being passed to a dedicated Verifier/Critic stage, aiming to improve reliability over conventional single-step RAG pipelines.

Key Features:

• Vector retrieval, knowledge graphs, SQL/API sources, and web search integrated into a unified research workflow
• Evidence-fusion layer that combines information from different sources before downstream reasoning
• Dedicated Verifier/Critic stage that checks fused evidence before final response generation
• System separated into distinct stages: reasoning, retrieval, evidence aggregation, verification, and final response generation

Exploration:

The project explores how multi-source evidence and explicit verification can improve reliability compared with conventional single-step RAG pipelines.

Technologies Used: Python, RAG, Vector DB, Knowledge Graph, SQL/API, Web Search

Project Overview:

A personalized memory architecture that moves beyond retrieving similar historical messages toward generating a compact representation of the user's current goals, preferences, important facts, and recent changes.

Key Ideas:

• Identifies limitations of retrieval-based memory: outdated information, contradictions, changing goals, and irrelevant historical context
• Proposes a memory-processing layer that interprets long-term interaction history and generates context optimized for downstream LLM reasoning
• Explores potential improvements in personalization, contradiction handling, context-window efficiency, and long-term conversational coherence
• Defines a research question investigating whether an understanding-oriented memory model can provide more useful contextual information than conventional retrieval-based memory

Status: Architecture / Research Prototype

Technologies Used: AI Systems Architecture, LLMs, Retrieval, Memory

Project Overview:

The Bio-RAG Foundation Agent is an advanced Retrieval-Augmented Generation system designed for biomedical document analysis. It combines state-of-the-art embedding models with a robust vector database to provide accurate, context-aware responses to complex scientific queries.

Key Features:

• Intelligent document ingestion with automatic chunking strategies
• Gemma2 embedding generation for semantic understanding
• Weaviate vector database for efficient similarity search
• Advanced retrieval with hybrid search (keyword + semantic)
• Context-aware answer generation with source attribution
• FastAPI-based REST API with comprehensive documentation
• Performance monitoring and evaluation dashboards

Technical Implementation:

Built using Python with FastAPI framework, the system leverages Ollama for local LLM deployment, Weaviate as the vector database, and custom embedding models. The architecture follows microservices principles with Docker containerization for easy deployment and scaling.

Challenges & Solutions:

• Challenge: Handling diverse document formats and structures
• Solution: Implemented flexible parsing pipeline with format-specific handlers
• Challenge: Optimizing retrieval accuracy while maintaining speed
• Solution: Hybrid search combining keyword matching with semantic similarity
• Challenge: Ensuring response quality and relevance
• Solution: Multi-stage filtering and ranking with confidence scoring

Technologies Used: Python, FastAPI, Ollama, Gemma2, Weaviate, Docker, Redis, PostgreSQL, Prometheus, Grafana

Project Overview:

A comprehensive restaurant recommendation platform that combines multiple recommendation approaches to provide personalized dining suggestions. The system analyzes user preferences, dining history, and restaurant characteristics to deliver accurate recommendations.

Key Features:

• Hybrid recommendation engine (collaborative + content-based)
• User profiling and preference learning
• Real-time recommendation generation
• Location-based filtering and suggestions
• Social features and friend recommendations
• Advanced search and filtering options
• User reviews and rating system

Technical Implementation:

The backend is built with Django REST API, utilizing TensorFlow for machine learning models and Scikit-learn for traditional ML algorithms. The system employs TF-IDF vectorization for content-based filtering and matrix factorization for collaborative filtering.

Algorithm Details:

• Collaborative Filtering: Matrix factorization with regularization
• Content-Based: TF-IDF with cosine similarity
• Hybrid Approach: Weighted combination of both methods
• Cold Start Problem: Content-based for new users

Technologies Used: Python, Django, TensorFlow, Scikit-learn, PostgreSQL, Docker, Redis, Nginx