🧾 “I Agree” Without Reading? Let GenAI Do It For You.

Ever clicked “I agree to the Terms & Conditions” without reading them?
You’re not alone. But what if a GenAI agent could read them for you… and explain it like a human?

👋 Meet the SaaS Terms Simplifier & Risk Analyzer Agent

My GenAI Capstone project turns walls of legal jargon into simple summaries and surfaces hidden red flags — all within an interactive, modular GenAI pipeline.

It’s not just a prototype — it’s a working AI legal assistant that:

• 📃 Reads any SaaS Terms or Privacy Policy

• ✍️ Summarizes them in plain English

• 🚩 Flags legal risks like forced arbitration or data selling

• 💬 Lets you chat with the document (RAG chatbot)

• 📤 Exports the insights in JSON, Markdown

---

🧠 The Problem: Legal Docs are Designed to Confuse

We’ve all signed up for a SaaS product — Zoom, Canva, Notion, you name it — and casually accepted their Terms of Service or Privacy Policy. But:

• What are we really agreeing to?

• Can they delete our data anytime?

• Are we giving them permission to sell our info?

These documents are often long, boring, and intentionally vague. And that’s a problem — both for users and companies.

---

💡 The Solution: AI-Powered Legal Translator

My idea? Build an AI agent that:

• Reads any SaaS Terms & Conditions or Privacy Policy

• Summarises it in simple language

• Flags risky clauses that need your attention

• Lets you chat with it like a legal assistant

• Outputs structured results in JSON, Markdown, PDF — your choice

And I didn’t want to stop at a basic proof-of-concept. I took it to an advanced level, using LangChain, Gemini, Retrieval-Augmented Generation (RAG), and IPython widgets for an interactive notebook experience.

---

🛠️ Tech Stack

• LangChain + Gemini: Agent orchestration and LLM capabilities

• IPython + ipywidgets: Interactive UI inside Kaggle Notebook

• Markdown / JSON / PDF: Output formats

• Pandas: Data handling

• FAISS: For embedding-based search in FAQ bot

• Prompt Engineering, Structured Output, RAG, Long Context: GenAI features

---

🧩 Core Features (Step-by-Step)

Let me walk you through what the agent actually does — step by step:

The code snippets provided are just for example

---

1️⃣ Upload the T&C File or Paste Text

The user uploads a PDF or pastes legal text directly. They also provide the SaaS app name (e.g., “Notion”). This acts as context for personalization.

Under the hood:

• The text is smartly chunked using a hybrid strategy (based on tokens and headings)

• Long documents are handled via recursive splitting

• Each chunk maintains coherence and context

def smart_split(text):
    # Recursive chunker using token + heading heuristics
    if len(text.split()) < 300:
        return [text]
    return re.split(r'\n[A-Z][^\n]{0,50}\n', text)

---

2️⃣ Plain English Summarization

Each chunk is processed by the Summarizer Agent, which:

• Rewrites legalese into human-friendly English

• Groups content into sections like "Privacy", "Payments", "Account Termination"

• Supports two user personas:

  • 🧑‍💻 For technical folks

  • 👶 For non-technical folks

🔍 Powered by:

• Prompt Engineering + Few-shot examples

• Gemini’s structured JSON output

prompt = PromptTemplate.from_template("""
Summarize this clause in plain English for a {persona}:

Clause:
{text}

Return as:
{
  "section": "...",
  "summary": "...",
  "impact": "..."
}
""")

---

3️⃣ Red Flag Detector

Next, a Red Flag Agent scans the document to detect:

• Risky language like “we may sell your data”

• Tricky clauses like forced arbitration or auto-renewals

• Assigns severity tags: 🟢 Safe | 🟡 Caution | 🔴 High Risk

The result?

• A clean, categorized list of red flags

• Color-coded highlights with tooltips for explanations

• JSON export of flagged clauses + reasoning

🔍 Powered by:

• Function Calling

• Grounding (clause + explanation)

• Structured outputs

def detect_red_flags(clause):
    if "we may terminate" in clause.lower():
        return {
            "risk_type": "Termination Clause",
            "severity": "🔴",
            "explanation": "The service can terminate your account at any time without notice."
        }

---

4️⃣ Executive Summary

Too busy for details? The agent gives a quick, crisp TL;DR:

• 3 versions available:

  • 🧠 Legal-focused

  • ✨ User-focused

  • 🔎 Executive summary for product teams

You’ll know in a glance:

“This T&C looks safe overall, but there’s a clause on auto-renewal that you might want to read.”

---

5️⃣ Ask Anything with the RAG Chatbot

Here’s where it gets cool.

You can ask questions directly, like:

• “Can they terminate my account at will?”

• “Do they collect personal health data?”

The chatbot uses:

• FAISS embeddings to search relevant chunks

• Gemini to generate focused answers grounded in the source

🔍 Powered by:

• Embeddings

• RAG

• Long context window

retriever = FAISS.load_local("terms_db").as_retriever()
qa_chain = RetrievalQA.from_chain_type(
    llm=llm,
    retriever=retriever,
    return_source_documents=True
)

---

6️⃣ Export Results

Choose your format:

• 📄 Markdown (for normal users)

• 🧾 JSON (for dev/legal teams)

---

⚙️ Codebase Overview

Here’s the modular architecture I designed:

saas_term_agent/
├── app.py                 # Main UI using ipywidgets
├── agents/
│   ├── summarizer.py      # Summarizer Agent
│   ├── red_flag.py        # Red Flag Detector
│   └── summary_gen.py     # Executive Summary Generator
├── utils/
│   ├── text_splitter.py   # Token-based and heading-based chunking
│   └── output_parser.py   # Export functions (Markdown, JSON, PDF)
├── prompts/
│   ├── summarizer.txt     # Prompt template with examples
│   └── red_flag.txt       # Red flag detection patterns

To keep the notebook clean, app.py is imported from a script file, but all logic and key components are demonstrated cell-by-cell in the notebook.

---

📈 What GenAI Capabilities Did I Use?

✅ Prompt Engineering
✅ Few-shot examples
✅ Function Calling (simulated via structured output)
✅ Structured output (Markdown + JSON)
✅ Grounding
✅ RAG (Retrieval Augmented Generation)
✅ Long Context Support

All requirements for the Capstone? ✅ Met and exceeded.

---

🔥 Challenges I Faced

• Streamlit is not supported in Kaggle, so I built the whole UI in IPython Widgets (with styling, progress bars, tabs, and toggles).

• Gemini’s output sometimes wasn’t perfectly structured — I had to build custom parsers and validators.

• Getting chunking right was non-trivial. Poor chunking = poor summarization.

---

🌟 Why This Project Stands Out

• ✅ It’s real-world relevant (we all click “I agree”)

• ✅ It’s not just a GenAI demo — it solves an actual user pain point

• ✅ It’s modular, scalable, and beautifully structured

• ✅ It combines multiple GenAI capabilities, not just one

• ✅ It has a strong UX layer, even inside a notebook!

---

⛔ Limitations & Future Scope

---

🔮 What’s Next?

Imagine this agent as:

• 🔌 A Chrome Extension for auto-analysis of Terms pages

• 🏢 A SaaS Procurement Tool for startups

• 🔁 A ToS Tracker that alerts you when terms change

• 💬 An API plugin for product onboarding

---

🏁 Final Thoughts

The SaaS Terms Simplifier & Risk Analyzer Agent isn’t just a capstone — it’s a launchpad.

It proves how GenAI can transform how we interact with boring-but-important legal docs. This is just the start — imagine this embedded in browsers, signup flows, or enterprise SaaS reviews.

✨ Let GenAI read the fine print for you — so you don’t have to.

Want Source Code?

click here

---

If you liked this project or want to collaborate, follow me @neurontist on Hashnode or connect on LinkedIn.

Let’s make legalese understandable — one clause at a time!