Developed a RAG engine to combine the power of large language models with a dynamic knowledge base of federal regulations.

• Integrated Pinecone as the vector database for storage and retrieval of document embeddings
• Utilized Pinecone's similarity search capabilities to quickly identify the most relevant regulatory sections
• Incorporated Cohere AI's reranking model to further refine search results
• Enhanced overall quality of responses using semantic understanding capabilities

Built a proof-of-concept system using LangGraph to generate and improve SQL queries through iterative prompt refinement.

Implemented three collaborative agents:

• Query Generator: Converts natural language to SQL using context-aware prompting
• Evaluator: Checks query structure and identifies common inefficiencies
• Optimizer: Refines prompts based on pattern recognition from successful queries

Demonstrated 15% improvement in query correctness compared to direct LLM generation on a test set of 50 common database operations.

Devised a FAQ generation system in PyTorch through the fine-tuning of Llama-2-7B and Llama-3-8B LLMs.

• Applied Quantized-Low Rank Adaptation technique (Q-LoRa) to efficiently fine-tune the models in resource-constrained environments
• Formulated a prompt-based learning strategy to enhance contextual understanding
• Benchmarked the fine-tuned Llama models against Google's Flan T5-Large and Meta's BART
• Attained a BERT Score of ~0.8, signifying high fidelity in semantic similarity compared to human-written answers