Regulation Repository
Terminology Repository
Feature Checks
Provide context & term mappings
Reconcilation & Learning

Regulo

Regulo is an AI-assisted platform designed to determine whether product features require geo-specific compliance logic. It converts dense regulatory texts into structured, actionable compliance insights, making complex regulations interpretable and traceable for global product teams.

Inspiration

Validating product features against regulations worldwide is extremely challenging due to the diversity of legal requirements and the complex jargon that often requires domain-specific knowledge. Teams struggle to ensure compliance because regulations are dense, vary by region, and can be difficult to interpret without expert knowledge.

We chose this track because we were inspired by the challenge of using Large Language Models (LLMs) to address such a critical and impactful problem faced by TikTok. In our solution, we aimed to leverage strong reasoning capabilities of existing models while applying strategies to minimize hallucinations. We focused on making regulatory evaluation scalable, accurate, and auditable, ensuring features are correctly flagged for geo-specific compliance needs.

Main Challenge Addressed

Regulo is designed to tackle several core challenges:

LLM Lack of Context: Ensuring the model has enough background knowledge to make informed evaluations across diverse regulations.
Unaudited LLM Decision Process & Source Referencing: Making the LLM's reasoning traceable with references to sources and justifications for compliance decisions.
Ambiguity in Acronyms, Jargon, and Domain-Specific Term: Resolving confusion arising from industry-specific terms and abbreviations.
Hallucinations Generated by the LLM: Minimizing incorrect or fabricated outputs while maintaining reasoning capabilities.
Human-in-the-Loop Intervention: Allowing human oversight to refine evaluations, provide additional context, and correct uncertain results, ensuring accountability and auditability.

Solution Overview

Regulation Repository & Ingestion Pipeline

Regulation Repository: The central, authoritative source of all regulatory knowledge in the system. All reasoning and references by the LLM are constrained to this repository, ensuring traceability.
Regulation Ingestion Pipeline: Uses AI agents to extract atomic units of regulation from uploaded documents, called Regulatory Obligation Units (ROUs).
MapReduce Optimization with StateGraph: Large documents are split into smaller chunks using a node-based MapReduce model with LangGraph's StateGraph. Each chunk is processed in parallel by an extraction agent, then deduplicated to remove overlapping ROUs.
Vector Database (Chroma): ROUs are also stored in a vector database for our RAG (Retrieval-Augmented Generation) tool. The AI agent can query the vector DB using text queries to retrieve relevant regulations efficiently, avoiding the need to loop through all stored regulations, which would be time-consuming.

Terminology Repository

Serves as the source of all mappings for acronyms, jargon, and domain-specific terms to their possible meanings.
Allows users to do CRUD operation on the mappings.
Queried by the AI agent whenever it needs to retrieve existing mappings for a specific term.
Users can still provide alternative mappings for a specific feature evaluation, which will be incorporated into the reasoning process.

Feature Evaluation

Two-Step Workflow: Utilizes two dedicated LangGraph ReAct agents with OpenAI GPT-5-nano for strong reasoning.
1. Terminology Mapping Agent: Scans the feature title and description, identifies acronyms and jargon, and determines the most appropriate mapping using both the feature context and the terminology repository.
2. Evaluation Agent: Receives mapped terminologies from the first agent and accesses the Regulation Repository. It understands the feature, reasons about all potentially relevant regulations, and crafts queries to retrieve applicable ROUs.
Outputs a flag (yes, no, unknown), detailed reasoning, and suggested next steps for users.
AI agents may request additional context for ambiguous feature descriptions. User can retrigger the evaluation after enriching the feature description with more context.
Users can also edit the assumed terminology mappings that aren't accurate enough to improve the evaluation accuracy.

Reconcilation & Learning

Users can reconcile the results provided by the AI agents to correct or refine evaluation outputs.
A learning agent compares the initial AI evaluation with the reconciled results and updates the system prompt of the evaluation agent if applicable, improving future reasoning and decision-making accuracy.
This approach ensures that human insight continuously enhances the AI while maintaining auditability and traceability.