QuickRef — Banking Knowledge Assistant

Leading the migration of a RAG and agentic banking/finance knowledge assistant off the Gemini API onto a self-hosted, fine-tuned open-source LLM.

Project Overview

QuickRef is QuickFox's RAG and agentic (LangGraph ReACT) banking/finance knowledge assistant. Its original retrieval and agentic pipeline — semantic chunking, Qdrant vector search, multi-turn tool use — was built by a QuickFox engineering predecessor. My contribution is migrating the assistant's underlying generation model off the Gemini API onto a self-hosted, LoRA/QLoRA fine-tuned open-source transformer model, targeted at Nepali/Devanagari financial-document Q&A, to eliminate per-call inference costs while preserving the existing RAG and agentic behavior.

Key Features

  • •LoRA/QLoRA fine-tuning + SFT of open-source transformer models
  • •Devanagari tokenizer adaptation to reduce token fragmentation on Nepali text
  • •Curated Nepali/Devanagari financial-domain training data
  • •Preserves the existing RAG and ReACT agentic workflow, benchmarked against the Gemini baseline

Technology Stack

  • •Hugging Face Transformers, PEFT, TRL
  • •LoRA/QLoRA fine-tuning, Python, PyTorch
  • •LangChain, LangGraph (ReACT Agents)
  • •Qdrant (vector search), Gemini API (previous backend)
← Back to Work
;