JarvisFi
Multilingual LLM finance assistant with RAG grounding and async job processing.
TL;DR
A conversational finance assistant that answers in English, Tamil, Hindi, or Telugu, grounded in retrieved market context via RAG over Hugging Face and IBM Watsonx models. Took 2nd place at a national paper presentation.
- Real-time Telemetry
- <200ms API Response (FastAPI + Redis Cache)
Problem
Financial guidance in India is English-first and jargon-heavy. Non-English speakers get either nothing or a generic chatbot that hallucinates numbers.
Approach
- 01
Built a language-detection front door routing English, Tamil, Hindi, and Telugu queries through one reasoning path.
- 02
Grounded answers with a RAG layer over market and document context using Hugging Face and IBM Watsonx models.
- 03
Offloaded ingestion and long-running analysis to Celery workers with Redis as broker and response cache.
- 04
Secured multi-user sessions with JWT auth and tuned prompts against a hand-labelled evaluation set.
Key architectural decision
Implemented Redis caching and Celery task queues to handle asynchronous PDF report generation and RAG retrieval without blocking primary user request threads.
Impact
- First Runner-Up, National Level Paper Presentation at Coimbatore Institute of Technology (NEXERA'26).
- Top 10 Innovator at GenAI Hackathon.
- RAG grounding plus Redis caching cut unsupported numeric claims and repeat-query latency.
What shipped
Multilingual routing
Detect → translate → retrieve → respond in the source language.
RAG grounding
Hugging Face + IBM Watsonx retrieval context injected per query.
Async pipeline
Celery workers with Redis broker and cache behind JWT-scoped sessions.
Stack
- Python
- Hugging Face
- IBM Watsonx
- RAG
- Redis
- Celery
- JWT
- Streamlit
- NLP