AI Investment Education Assistant
Retrieval-Augmented Generation (RAG) for Financial Education
Case Study
AI assistant that helped new investors get reliable answers without crossing into financial advice
I designed and built a retrieval-augmented generation (RAG) assistant for an investment platform that answered educational questions using grounded source material and compliance guardrails.
Role
Solo developer
Domain
FinTech · Investment Education
Technologies
Python · Gemini · FAISS · FastAPI
Duration
Jun 2024 – Apr 2026
Project Snapshot
Problem
New investors often asked repetitive onboarding and financial education questions, but a raw LLM could hallucinate or accidentally provide personalized advice.
Solution
I designed a RAG assistant that retrieved approved content before generating an answer, keeping responses grounded in verified source material.
Impact
The system indexed 100+ financial education documents, connected retrieval and generation into one workflow, and helped reduce onboarding response time by 2 days.
Overview
Financial platforms receive many repeated questions from new investors, but answering them with a general-purpose LLM creates risk. The model may sound confident while giving unsupported, inaccurate, or overly personalized financial guidance.
This project was designed to solve that by routing every user question through a retrieval pipeline before generation. Instead of relying only on the model’s training data, the assistant searched a curated knowledge base, assembled relevant context, and generated responses grounded in approved educational material.
The Challenge
Hallucination risk
LLMs can generate fluent answers even when they do not have the right information. In a financial education setting, that creates trust and accuracy concerns.
Compliance sensitivity
The assistant needed to explain investment concepts clearly without crossing into personalized financial advice.
Onboarding friction
New users often asked the same questions repeatedly, creating a need for a scalable assistant that could handle common education and setup questions.
Fragmented documentation
Relevant information was spread across financial education pages, FAQs, and policy-style content, making it hard for users to find direct answers.
System Architecture
Rather than sending user questions directly to an LLM, the system routed every request through a retrieval pipeline before generation.
User Question
User asked an investment education question through WhatsApp.
Embedding
The query was converted into a vector embedding.
Retrieval
FAISS searched for the most relevant knowledge base chunks.
Context
Retrieved documents were inserted into a structured prompt.
Generation
Gemini generated an answer using the retrieved context.
Safe Response
The answer was checked against compliance rules before returning.
Compliance Guardrails
The assistant was designed to refuse personalized investment advice, avoid unsupported claims, and keep responses focused on financial education rather than recommendations.
End-to-end RAG workflow with retrieval, generation, and compliance filtering.
Technical Stack
- Python
- Gemini API
- FAISS
- FastAPI
- WhatsApp API
- RAG
- Vector Search
- Prompt Engineering
- LLMs
- SendGrid
What I Built
Document Ingestion Pipeline
Chunked and embedded 100+ financial education documents into a FAISS vector store, balancing chunk size, retrieval precision, and context length.
Conversation Engine
Built a multi-turn conversation engine with structured prompt templates, retrieval-augmented generation, and conversation history management.
Compliance Layer
Designed prompt and output guardrails to detect advice-seeking questions, enforce source-grounded generation, and return safe refusals when needed.
Backend Integration
Built a FastAPI backend to handle WhatsApp webhook events and connect onboarding workflows for new platform users.
Impact
Financial education documents indexed
Reduction in onboarding response time
Core pipeline components: embeddings, retrieval, prompting, LLM, guardrails, WhatsApp
What I Learned
Building this project taught me that strong AI products depend on much more than the model itself. The retrieval layer had the biggest impact on answer quality: chunk size, embedding choice, similarity thresholds, and context formatting shaped whether the assistant could answer accurately.
I also learned that compliance is both a technical and product design problem. The system needed clear rules for when to answer, when to refuse, and how to keep educational responses helpful without becoming personalized financial advice.