Geo-Regulator

Compliance_Analyzer_SingleFeature
Compliance_Analyzer_CSV
Compliance_Results

Inspiration

Global platforms like TikTok must launch features across multiple regions, each with its own evolving laws (EU DSA, California SB976, Utah Social Media Act, etc.). Today, compliance is mostly manual, slow, and expensive - companies risk regulatory fines, lawsuits, and damaged trust when geo-specific rules are missed. Hence, we wanted to build a tool that makes compliance proactive, fast, and transparent instead of being a blocker.

What it does

GeoRegulator is an AI-powered compliance analyzer that:

Takes in product feature descriptions (single or CSV bulk upload).
Uses RAG (Retrieval-Augmented Generation) to match features against region-specific regulations.
Flags whether a feature needs geo-specific compliance logic: REQUIRED, NOT_REQUIRED, or UNCERTAIN.
Provides reasoning, confidence scores, and direct citations from regulation text.
Highlights low-confidence results for human review, with editable reasoning and reviewer checkboxes.
Exports audit-ready CSVs that compliance teams can use directly.

How we built it

Backend: Python + Flask for serving the app.
LLM Pipeline: OpenAI (GPT-4o, GPT-3.5) and Google Gemini (gemini-2.5-flash).
RAG Layer: Regulation text files loaded from a /regulations directory and filtered by region.
Data Handling: Pandas for CSV processing, terminology mapping for TikTok’s internal jargon.
Frontend: HTML + TailwindCSS + vanilla JavaScript for the web interface (CSV upload & single feature mode).
Infrastructure: Dockerized and deployed on Google Cloud (Cloud Run).

Challenges we ran into

Parsing and structuring diverse regulation texts into a usable format.
Getting the LLM to return consistent JSON outputs (sometimes hallucinated extra text). = Mapping TikTok-style jargon (e.g., PF, ASL, CDS) into clear descriptions the model can reason about.
Handling performance trade-offs between accuracy vs. speed across OpenAI and Gemini models.
Building an interface that balances automation with human-in-the-loop review.

Accomplishments that we're proud of

Successfully implemented a RAG-powered compliance checker that cites the exact regulation text.
Built a dual-mode frontend (bulk CSV + single feature) that is clean, intuitive, and responsive.
Added human-in-the-loop review tools (editable reasoning, checkboxes, row highlighting).
Deployed a fully containerized solution on Google Cloud.

What we learned

How to integrate multiple LLM providers (OpenAI + Gemini) into a single pipeline.
The importance of structured prompting to force consistent, audit-ready outputs.
That compliance is not just about detecting risks — it’s about explaining them transparently.
How to build user trust by combining automation with human-in-the-loop intervention.
A ton about cloud deployment, Docker, and scaling lightweight AI apps.

What's next for Geo-Regulator

Add policy ingestion: upload PDFs of new laws and auto-integrate them into the regulation base.
Expand coverage to more global regulations (India IT Rules, UK Online Safety Bill, etc.).
Build a dashboard for compliance teams with trend analysis, risk heatmaps, and export to legal workflows.