An AI banking companion that explains UPI, NEFT, EMI, KYC & more in simple words, and helps you spot scam messages before you click.
🎬 Demo • ✨ Features • 👥 Who It's For • 🧠 Tech • ⚡ Quick Start • 🔒 Safety
A ~1 minute 45 second walkthrough covering the problem, who the app is for, a live tour of all three tools, how it works under the hood, and why you would use it.
Click the image to play the video.
 💬 Chat home Welcome message, one-tap example questions, language & detail settings |
 🗣️ Hinglish answers Ask in Hinglish, get a reply in Hinglish with a real-life example |
 📖 Banking Term Explainer Meaning, Hindi translation, example, steps, safety tip & related terms |
 🛡️ Scam Message Checker Risk level, red flags and exactly what to do next |
BankSaathi AI is a multilingual, AI-powered banking information assistant. "Saathi" (साथी) means companion, and that is the idea: a friendly guide that explains everyday banking in the language you are most comfortable with.
Banking is full of jargon (NEFT, RTGS, IMPS, IFSC, KYC, EMI…), most explanations are written in formal English, and fake "KYC update" or "account blocked" messages are a constant threat. BankSaathi AI tackles all three problems in one simple app.
⚠️ BankSaathi AI never touches your bank account. It gives general information only, so it cannot check balances, move money or block cards.
| Who | What they struggle with | How BankSaathi AI helps | |
|---|---|---|---|
| 🎓 | Students & first-time account holders | New to UPI, debit cards, statements, IFSC | Plain-language explanations with real-life examples |
| 👨👩👧 | Families & senior citizens | Fear of making a mistake or being tricked | Patient follow-up questions + a scam checker for suspicious messages |
| 🗣️ | Hindi & Hinglish speakers | Most banking content is formal English | Ask and receive answers in हिन्दी, English or Hinglish |
| 🛒 | Small shop owners using UPI | Understanding payment flows and PIN safety | Clear explanations of UPI, NEFT/RTGS/IMPS and safe practices |
| 📱 | Anyone who just got a "suspicious" SMS | "Is this real or a scam?" | Paste it in, get a risk level, red flags and next steps |
- 🗣️ Speak naturally: ask in Hindi, English or Hinglish and get a reply in the same language.
- 🧒 Simple or detailed: switch between a short, beginner-friendly answer and a step-by-step one.
- 🛡️ Spot scams early: a dedicated checker explains why a message looks risky.
- 🔐 Private by design: OTPs, PINs, CVVs and long card numbers are blocked before they reach the AI.
- 🎯 Honest about limits: the AI is instructed not to invent fees, limits or interest rates, and to point you to your bank, RBI or NPCI instead.
- 🚪 Zero friction: no sign-up, no bank login. Just ask.
| Page | What it does | Highlights |
|---|---|---|
| 💬 Chat with BankSaathi | Multilingual banking chatbot with conversation memory | Reply-language selector (Auto / Hindi / Hinglish / English) • Simple vs Detailed mode • One-tap example questions • Clear-chat button |
| 📖 Banking Term Explainer | Type a term (EMI, KYC, IFSC…) and get a structured explanation | English + Hindi meaning • Real-life example • How it works (steps) • Safety tip • Related terms • Raw JSON view |
| 🛡️ Scam Message Checker | Paste a suspicious SMS / WhatsApp / email | Low / Medium / High risk badge • Scam type • Red flags • "What you should do" checklist • Cyber-crime helpline reminder (1930) |
flowchart LR
U([👤 User]) --> UI[🖥️ Streamlit UI<br/>3 pages + sidebar settings]
UI --> G{{🔐 Privacy Guard<br/>OTP / PIN / CVV / card no.}}
G -- "blocked ⚠️" --> W[Warning shown<br/>nothing sent to AI]
G -- "clean ✅" --> R{Which page?}
R -- Chat --> C[💬 Dynamic system prompt<br/>+ last 12 messages]
R -- "Term / Scam" --> P[🧩 ChatPromptTemplate<br/>+ format instructions]
C --> L1[(🤖 ChatGroq<br/>temperature 0.3)]
P --> L2[(🤖 ChatGroq<br/>temperature 0)]
L2 --> PA[📦 PydanticOutputParser<br/>BankingTerm / ScamCheck]
L1 --> OUT([✅ Answer in UI])
PA --> OUT
style U fill:#FF9933,color:#1B1F3B,stroke:#FF9933
style UI fill:#1F2A7A,color:#fff,stroke:#1F2A7A
style G fill:#E92063,color:#fff,stroke:#E92063
style W fill:#FFE3E3,color:#8a1f1f,stroke:#E92063
style C fill:#EEF0F8,color:#1B1F3B,stroke:#1F2A7A
style P fill:#EEF0F8,color:#1B1F3B,stroke:#1F2A7A
style L1 fill:#F55036,color:#fff,stroke:#F55036
style L2 fill:#F55036,color:#fff,stroke:#F55036
style PA fill:#8A2BE2,color:#fff,stroke:#8A2BE2
style OUT fill:#138808,color:#fff,stroke:#138808
| Layer | Technology | Used for |
|---|---|---|
| 🐍 Language | Python 3.10+ | Application logic |
| 🖥️ UI | Streamlit | Chat UI, multi-page sidebar navigation, session state, themed via .streamlit/config.toml |
| 🔗 Orchestration | LangChain (langchain-core, langchain-groq) |
Prompt templates, message roles, model wrapper, output parsing |
| ⚡ LLM inference | Groq · model openai/gpt-oss-20b |
Fast hosted inference |
| 📦 Validation | Pydantic | Typed schemas for structured AI output |
| 🔑 Config | python-dotenv | Loading GROQ_API_KEY from .env |
| Concept | Where / how it is used |
|---|---|
| Chat models & sampling control | ChatGroq is created twice: temperature 0.3 for natural but focused chat, temperature 0 for predictable structured extraction. Both are cached with @st.cache_resource. |
| Dynamic system prompts | build_system_prompt() rebuilds the SystemMessage on every call from the sidebar choices (language + detail level), so setting changes apply instantly. |
| Message roles & short-term memory | Conversation is stored as HumanMessage / AIMessage objects in st.session_state; the system message is added fresh each turn. |
| Sliding-window memory | Only the last MAX_HISTORY_MESSAGES = 12 messages are sent to the model, which keeps context (and cost/latency) bounded. |
| Prompt templates | ChatPromptTemplate with {term}, {message} and {format_instructions} placeholders for the two structured tools. |
| Structured output | PydanticOutputParser converts the LLM reply into validated BankingTerm and ScamCheck objects, which drive the UI (badges, lists, JSON view). |
| Guard-rail prompting | Prompts forbid inventing fees/limits/rates, require "unsure" over guessing, restrict the bot to banking topics and always route fraud to 1930 / cybercrime.gov.in. |
| Privacy guard (pre-LLM filter) | contains_sensitive_info() blocks a message before any API call when it has a secret word (OTP, PIN, CVV, password, passcode) plus 4+ digits, or any run of 12+ digits (card / Aadhaar-length numbers). It is a simple heuristic, so treat it as a safety net, not a guarantee. |
| Graceful error handling | Failed calls remove the unanswered question from history and show a friendly error instead of crashing. |
BankSaathi AI/
├── 📁 app/
│ └── app.py # Entire Streamlit app (UI + LLM logic)
├── 📁 assests/
│ ├── logo.jpg # Full logo (README)
│ └── logo_small.jpg # Sidebar / page icon
├── 📁 preview/
│ ├── 01-chat-home.png # Screenshots
│ ├── 02-chat-hinglish.png
│ ├── 03-term-explainer.png
│ ├── 04-scam-checker.png
│ ├── banksaathi-demo.mp4 # Product demo video
│ └── video-thumbnail.png # Poster image used in this README
├── 📁 .streamlit/
│ └── config.toml # Theme (indigo + cream)
├── .env.example # Copy to .env and add your key
├── .gitignore
├── requirements.txt
├── What is BankSaathi AI.md # Plain-language project description
└── README.md
Prerequisites: Python 3.10+ and a free Groq API key.
# 1. Clone the repository
git clone https://github.com/Rohit-coder-py/BankSaathi-AI.git
cd BankSaathi-AI
# 2. (Recommended) create a virtual environment
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
# 3. Install dependencies
pip install -r requirements.txt
# 4. Add your API key
cp .env.example .env # Windows: copy .env.example .env
# then open .env and set GROQ_API_KEY="your_real_key"
# 5. Run the app (from the project folder)
streamlit run app/app.pyThe app opens at http://localhost:8501 🎉
- Push this repo to GitHub (never commit
.env; it is already in.gitignore). - On share.streamlit.io, create a new app and set the main file path to
app/app.py. - In Advanced settings → Secrets, add
GROQ_API_KEY = "your_real_key". - Deploy, then paste the live URL into the
your-app-nameplaceholders at the top of this README.
Warning
Never share your OTP, UPI PIN, ATM PIN, password, CVV or full card number in any chat, including this one.
- 🚫 The app does not access bank accounts, check balances, verify transactions, move money or block cards.
- 🔐 Messages that look like they contain secrets are stopped before they are sent to the AI.
- 📉 Fees, limits, timings and interest rates differ between banks, so the assistant tells you to confirm with your bank's official app/website, RBI or NPCI.
- 🧠 AI can make mistakes. Always verify important information through official channels.
- 🚨 Lost money or shared your details? Call your bank immediately and report on the National Cyber Crime Helpline 1930 or at cybercrime.gov.in.
- 🌐 More Indian languages (Marathi, Tamil, Bengali…)
- 🎙️ Voice input for hands-free questions
- 📚 Retrieval-augmented answers from official RBI / NPCI documents
- 🧪 Automated tests for the privacy guard and parsers
- 📱 Mobile-friendly layout polish
Rohit Jha: building AI/ML and GenAI projects.
⭐ If this project helped you, please consider giving it a star! ⭐
BankSaathi AI is an educational / portfolio project that provides general banking information only. It is not an official service of any bank, is not affiliated with RBI or NPCI, and does not offer financial, legal or investment advice.
BankSaathi AI: Your banking companion, in your language.