Legal Lens Logo
Here is our batch processing results for 7 samples
RAG workflow
ChromaDB building workflow

Legal Lens – TikTok Geo-Compliance Automation System

TikTok TechJam 2025 – Track 3 Submission by Team Durian Men

📌 Features and Functionality

Legal Lens is an automated system for determining geo-specific compliance requirements of TikTok features using AI-powered legal analysis.
It transforms compliance detection from manual guesswork into traceable, auditable outputs through:

Core Features

Automated Compliance Detection – Analyzes feature descriptions and determines if they require geo-specific legal compliance
Semantic Legal Search – Uses AI embeddings to find relevant legal articles from a comprehensive knowledge base
TikTok Jargon Translation – Automatically expands 50+ internal terminology terms (ASL, GH, T5, etc.) for proper legal context
Confidence Scoring – Provides reliability metrics and legal citations for compliance determinations
Batch Processing – Analyzes multiple features simultaneously via CSV input/output with simple rate limiting
Audit Trail Generation – Creates detailed reasoning and legal references for every decision

🏗️ System Architecture

Legal Document Processing – Article-based chunking of 5 regulatory frameworks into multiple compliance-focused segments
Vector Embeddings – BAAI/bge-small-en-v1.5 model generates semantic embeddings for legal text retrieval
ChromaDB Vector Database – Persistent storage enabling sub-second semantic search across legal corpus
Provence AI Integration – Prunes legal documents to extract only compliance-relevant obligations and requirements
Structured JSON Pipeline – Machine-readable outputs with verdicts, reasoning, and legal citations

⚡ Performance

Fast Processing – sub 7 seconds per feature analysis (vs 15–30 minutes manual legal review)
98% Time Savings – Dramatically reduces manual legal review overhead
Multi-Jurisdiction Coverage – Supports 5 major regulatory frameworks simultaneously

💵 Cost Efficiency

Low API Cost – Average input/output API call cost to Sonar was below 1 cent per request
Affordable at Scale – Enables batch analysis of hundreds of features with negligible expense
Predictable Pricing – Fixed cost per request ensures transparent budgeting

🛠️ Development Tools Used

Python 3.8+ – Primary programming language
Virtual Environment – Dependency isolation and management
Git – Version control and collaboration
Visual Studio Code – Main development environment

🔌 APIs Used in the Project

Perplexity API – Sonar model integration for AI-powered legal analysis and compliance determination
FastEmbed API – BAAI/bge-small-en-v1.5 embedding model for high-performance semantic search
Provence AI API – Content pruning service for compliance-relevant legal text extraction

🧩 Design Decision: Direct API vs LangChain

We chose direct HTTP calls to Perplexity over LangChain for simplicity and control.

✅ Transparency in outputs
✅ Minimal dependencies
✅ Direct error handling and response parsing
❌ LangChain abstraction unnecessary for our single-call use case

📚 Libraries Used

ChromaDB – Vector database for persistent semantic search
FastEmbed – High-performance embeddings
Requests – HTTP client for API interactions
Python-dotenv – Environment variable management
Transformers – Hugging Face library for Provence AI integration
NLTK – Text preprocessing for Provence integration
NumPy – Vector operations and embedding manipulation

⚖️ Relevant Problem Statement

From Guesswork to Governance: Automating Geo-Regulation with LLM

As TikTok operates globally, every product feature must dynamically satisfy dozens of geographic regulations (e.g., Brazil’s data localization, GDPR).

Without automation, TikTok faces:

⚖️ Legal exposure from undetected compliance gaps
🛑 Reactive firefighting when regulators inquire
🚧 Manual overhead in scaling global rollouts

✅ System Benefits

Proactive Legal Guardrails
- Flags compliance needs during feature development
- Provides confidence scoring + legal citations
- Escalates to human intervention if uncertain
Audit-Ready Evidence
- Structured JSON outputs with reasoning, references, and batch logs
Regulatory Traceability
- Persistent compliance records with citations and confidence metrics

📂 Assets Used

Legal Document Corpus – 5 major regulatory frameworks
TikTok Jargon Dictionary – 50+ terms mapped (ASL, GH, T5, Spanner, Jellybean, Snowcap, etc.)
Feature Dataset – 35+ test features for validation
Pre-trained Embedding Models – BAAI/bge-small-en-v1.5 (~1GB)

📈 Additional Datasets Created

Enhanced TikTok Jargon Dictionary

TikTok-specific abbreviations (e.g., ASL = Age-sensitive logic)
System codenames (e.g., Spanner = rule engine, Jellybean = parental controls)
Operational terms (e.g., ShadowMode, EchoTrace, Redline)
Compliance/security terms (e.g., GDPR, CCPA, PII, MFA, RBAC)
Ambiguity handling with multi-meaning resolution

Compliance-Focused Legal Embeddings

Extracted only obligations & compliance requirements
Removed irrelevant definitions/background info using Provence
Produced 42 focused legal segments

Smart Jargon Resolution System

Provide all meanings (TikTok + standard definitions)
Context-aware analysis by Perplexity Sonar
Coverage across 50+ terms
Support for multi-meaning abbreviations

🎯 Task Requirements Fulfillment

🛡️ Mitigate Regulatory Exposure

Proactive detection of compliance needs
Multi-jurisdiction coverage (Utah, California, Florida, Federal, EU)
Confidence scoring + legal citations
Human-in-the-loop design via “uncertain” verdicts

📋 Enable Audit-Ready Transparency

Structured JSON with verdict + reasoning + references
Batch processing logs for evidence
Immediate response to audits

💰 Reduce Compliance Governance Costs

98% time savings (~7s vs 30min per feature)
Automated batch processing for hundreds of features
Eliminates manual research
Ensures consistency + scalability

🧗 Challenges Faced

Embedding Legal Texts for Optimal Retrieval
Finding the right technique to embed dense legal language was difficult. We had to test out multiple chunking and embedding techniques so that ChromaDB retrieval returned just the right amount of context, too small lost meaning, too large returned irrelevant material.
Prompt Engineering for Sonar
Due to the independence of each API call to Sonar, we had to continuously re-engineer and refine the single-shot prompt to ensure one query reliably produced the desired structured JSON verdicts, reasoning, and citations without requiring multiple retries.

⚠️ Limitation

This is a prototype demonstrating feasibility.
Production deployment would require further validation and legal review.

👥 Team Info

Team: Durian Men
Track: Track 3 – From Guesswork to Governance: Automating Geo-Regulation with LLM
Focus: Boosting LLM Precision & Full Automation for geo-compliance detection

This system transforms TikTok’s compliance process from reactive firefighting to proactive legal guardrails, enabling audit-ready transparency for global feature rollouts.