Frontpage #1
Frontpage #2
Privacy policy entering
Loading page...
Result page #1
Result page #2
Comparison of two companies

Inspiration

FinePrint: AI Legal Risk Detector

About The Project

You've seen the notification: "We updated our Terms of Service" or "This is our privacy policy." When you open it to take a glance, you realize that the average ToS is 5,000+ words. At 250 words/minute, that's 20+ minutes to read.

In this fast-moving world, nobody has time to read thousands of words or large numbers of pages. Hence, we ignore it or just agree to it with a 5-second read.

What we don't realize is that with a single click of "agree" to the terms of companies or organizations, we are essentially selling our own private data around the world.

According to a 2024 study, 68% of Terms of Service contain clauses allowing companies to sell or share your personal data with third parties without explicit consent. Meanwhile, data breaches expose millions of records annually—often because of overly broad data access clauses buried in fine print.

But here's the real question: Who has time to read a 500-page document? Nobody. And every application and organization has privacy policies with hundreds of potential risks. So do we just stop using them?

That's where FinePrint steps in.

FinePrint is an AI-powered application that allows you to paste any legal document and analyzes it instantly. It can read those 500+ pages of legal documents and contracts, surface all the risks involved, and rate how critical they are. It provides users with different defense strategies to avoid getting their data breached. It also features a comparison tool that lets you compare legal documents from multiple organizations and see which is safest to use.

Inspiration

The idea of creating FinePrint came from one of the applications I use in my daily life. It used to change its privacy policies frequently, and whenever I tried to read them, they had millions of words and hundreds of pages.

I realized: Do companies make their policies and terms of use deliberately lengthy to discourage reading? I believe they do.

This realization motivated me to create an application that could help not just me, but millions of people facing the same problem—making legal documents understandable without requiring a law degree or hiring an attorney.

My Learnings

During the creation and refinement of FinePrint, I learned several critical lessons:

People Don't Read Legal Documents—By Design

People don't skip legal documents because they're lazy, but because they're designed to be unreadable. This is intentional. Most people don't realize how much data they're giving away with a single "Accept" button. While any AI chatbot can identify critical problems, FinePrint goes further—it provides solutions alongside the risks.

Model Selection Matters

Different LLMs have different strengths:

Llama 70B → More nuanced but slower
Qwen 2.5-72B → Faster, nearly as detailed
Qwen 2.5-7B → Smallest, still functional

Getting models to output valid JSON consistently requires careful prompt design. When one model fails (cold start, rate limit, timeout), having a backup strategy is essential to keep your app reliable.

Backend API Design & Frontend Communication

How to structure the /api/analyze endpoint and handle errors gracefully
Returning consistent JSON responses
Frontend-backend communication via Fetch API
Error handling, timeouts, and AbortController for cancellations

How I Built FinePrint

Architecture Overview

I chose a full-stack web approach: a Flask backend handling AI analysis, and a React frontend providing a beautiful and responsive interface. This separation allowed me to iterate on the UI independently while experimenting with different LLM models on the backend.

Backend: Smart Model Fallback Chain

The heart of FinePrint is a 3-model fallback chain using Hugging Face's Router API:

Llama 3.3-70B — For nuanced legal analysis (most capable)
Qwen 2.5-72B:fastest — Backup (nearly as good, faster)
Qwen 2.5-7B — Final fallback (smaller, but functional)

Request Flow:

User Input → Flask API → Model Fallback Chain → JSON Response → React Frontend

The API accepts raw document text and returns structured JSON:

Risk Score (0–100 scale)
TL;DR Summary (one-sentence catchline)
Red Flags Array (with severity levels and defense strategies)

To handle LLM output unreliability, I implemented three-tier JSON parsing:

Parse directly
Strip markdown fences
Regex extraction as last resort

Frontend: React + Vite + Tailwind CSS

I built the UI in React with Vite and Tailwind CSS because I needed:

Fast development cycles with hot module reloading
A modern dark-theme aesthetic
Legal documents are intimidating—the UI needed to feel approachable and engaging

Performance & Reliability

FinePrint analyzes documents in under 30 seconds with high reliability, even under load. The combination of:

Smart fallbacks
Honest error communication
Thoughtful UX design

...makes legal document analysis accessible to anyone—no law degree required.

Challenges While Building

Challenge 1: Cold Starts & Rate Limiting

Problem: Hugging Face serverless endpoints return 503 errors during model warm-up. Free tier has rate limits, making users wait frustratingly long.

Solution: I built a 3-model fallback chain + exponential backoff retry logic. If Llama times out, automatically try Qwen 72B, then Qwen 7B.

Result: Near-100% uptime with minimal user-facing delays.

Challenge 2: Testing with Real Legal Documents

Problem: Hard to test with diverse legal documents. Dummy data doesn't work—legal language is complex and specific.

Solution: I manually tested with real ToS from major companies (Amazon, TikTok, etc.), found patterns, and built test cases around them.

Result: Robust analysis across different document types and industries.

Challenge 3: Confidence Scoring Without False Certainty

Problem: How do you tell users "This is risky" without claiming false certainty? One wrong score can mislead someone into signing a bad contract.

Solution: I created a heuristic-based confidence score capped at 97%, never 100%. It combines:

Flag count
Severity levels
Document length

This honest approach actually builds more trust—users appreciate candor about uncertainty.

Challenge 4: Making the UI Engaging

Problem: Simply telling users "This is highly risky" or "This is low risk" made the UI feel boring and text-heavy.

Solution: I improved the UI with color-coded risk levels (red/amber/green) for quick visual scanning. Combined with detailed explanations, this makes risk assessment both clear and compelling.

Conclusion

When you read about FinePrint, you might ask: "Why not just use any AI/LLM to do the same thing?"

The answer lies in prompt engineering. Generating detailed results, comparing legal documents, categorizing risk levels, and determining critical thresholds requires specialized, time-intensive prompt design. This is where FinePrint excels—it's not just an interface to an LLM, it's a purpose-built consumer protection tool.

FinePrint delivers exactly what users need: