Polaris Trafficking Compass - Agentic AI Decision Platform

Technologies

python fastapi nextjs anthropic langgraph rag postgres pgvector mlops devops causal-inference kubernetes terraform gcp kustomize vault

Key Highlights

▸Principal ownership of an agentic LLM platform for evidence-based anti-trafficking insights
▸Designed multi-agent architecture combining predictive, causal, and retrieval-based reasoning
▸Integrated large-scale public, nonprofit, and research datasets into a unified AI system
▸Built grounded Q&A workflows to support policy, prevention, and investment decisions
▸Implemented LLM evaluation agents to continuously improve answer quality and reliability

Overview

Agentic AI platform designed to generate actionable, evidence-backed insights to combat human trafficking. Polaris Trafficking Compass distills complex, multi-source data into decision-ready knowledge that supports prevention strategies, law enforcement investment, victim services, and research-driven policymaking.

Architecture

Multi-agent, retrieval-grounded AI system:

Data Layer: Longitudinal Polaris trafficking data, socioeconomic indicators, and research literature
AI Layer: LangGraph-based agent orchestration with specialized retrieval and reasoning agents
Knowledge Store: PostgreSQL with pgVector for evidence-grounded RAG
Interface: Conversational agent paired with an analytical dashboard

Key Technical Contributions

Agentic Q&A System

Designed a conversational, multi-agent workflow that decomposes complex user questions into targeted retrieval and reasoning steps. Subagents collaboratively query trafficking data, socioeconomic indicators, and research evidence to produce coherent, grounded responses.

Predictive & Causal Modeling

Integrated predictive models to identify high-risk and low-risk counties alongside causal models to surface actionable levers (e.g., the impact of child poverty reduction on trafficking rates). Enabled users to move from correlation to causal understanding.

Evidence-Grounded RAG

Implemented a Retrieval-Augmented Generation pipeline using pgVector to ensure all LLM outputs are traceable to underlying evidence. Reduced hallucination risk and increased trust for policy and investment decisions.

LLM Evaluation & Quality Control

Built an LLM-based evaluation agent to score answer relevance, grounding, and clarity. Established a feedback loop to iteratively improve system responses and maintain high-quality outputs in sensitive, high-stakes domains.

Technologies

LLMs & Agents: LangGraph, OpenAI-compatible models
RAG: PostgreSQL with pgVector
Modeling: Predictive risk modeling, causal inference
MLOps: Evaluation agents, monitoring, iteration workflows
Backend: Python-based AI services

Impact

Enabled Polaris to explore trafficking dynamics through a unified, evidence-grounded AI system. Supports smarter allocation of resources, stronger policy recommendations, and research-driven interventions by transforming fragmented data into actionable intelligence.