Weekend POC · Open Source

🏦 Banking AI
Document Assistant

Ask any question about 1,000 banking documents in plain English. Disputes, complaints, statements and maintenance cases — all instantly searchable.

🚀 Try Live Demo 🏗️ Architecture 🔀 Interactions ⭐ GitHub
1,000
Banking Documents
3,948
Vector Chunks
256-dim
Titan Embeddings
~$1/mo
Running Cost
2
Query Engines
eStatements
400 documents
Disputes
250 documents
Complaints
200 documents
Maintenance
150 documents
$0.001
Cost per question
<2s
Response time
Explore the Project
Architecture & Design

Detailed technical diagrams showing how every component connects — built for engineers and architects.

Solution Architecture
🏗️
Architecture Diagram

Dual-engine design with Smart Query Router, real-time ingestion pipeline, and production migration path.

ClickHouse NL→SQL ChromaDB RAG Query Router
View Architecture→
Service Interactions
🔀
Interaction Diagram

Service-to-service API calls, dual-write pipeline flow, and the query routing decision tree.

Dual-write pipeline AWS Bedrock APIs ReplacingMergeTree
View Interactions→
Live Demo · Password Protected
🚀
Live Chat Demo

Try the assistant live. Ask questions in plain English — see ClickHouse aggregation and ChromaDB RAG in action.

🔐 DM for password Free to use
Open Demo ↗→
How It Works
Smart Query Routing

Every question is automatically classified and routed to the right engine — no manual switching needed.

👤
Banker
Plain English question
→
🔀
Query Router
Intent detection
→ →
📊
ClickHouse NL→SQL
Counts · trends · breakdowns
📦
ChromaDB RAG
Semantic · summarise · explain
→
🤖
Nova Lite LLM
Formats the answer
→
💬
Streamlit UI
Answer + citations
Example Questions
What Can You Ask?

Two types of questions, two engines — both answerable in plain English.

📊 Aggregation Questions → ClickHouse
"How many complaints were raised each year?"
"Which branch had the most disputes?"
"Total compensation paid per relationship manager?"
"Cases referred to CFPB by year?"
🔍 Content Questions → ChromaDB RAG
"Summarise Mathew Little's complaint"
"Why was dispute DSP00047 resolved against the customer?"
"Show all high priority complaints from Leeds"
"What did the customer claim in this case?"
Technology
Full Tech Stack

100% cloud-native, free-tier friendly, production-ready swap path.

☁️ AWS
AWS S3 AWS Bedrock Amazon Nova Lite LLM Titan Embeddings v2 (256-dim) AWS Textract OCR
🗄️ Data
ClickHouse Cloud ReplacingMergeTree NL→SQL Pipeline ChromaDB Cosine Similarity
🐍 Dev
LangChain Python 3.9 Streamlit Cloud pdfplumber python-dotenv
🏭 Prod
IBM FileNet P8 pgvector on RDS Claude Haiku IBM watsonx (optional) FastAPI + React
D

Dinesh Singh Panwar

Senior Technology Leader  ·  AI & Data Architecture  ·  Banking & Financial Services

Building practical AI solutions for regulated industries. This POC demonstrates how modern AI (RAG, NL→SQL, vector search) can unlock value from document repositories like IBM FileNet — at near-zero cost on AWS free tier.

AI & ML Banking Tech AWS Bedrock RAG Architecture IBM FileNet
LinkedIn Profile GitHub Repo