Skip to content

TrustGraph AI

TrustGraph AI uncovers coordinated marketplace abuse by connecting hidden account relationships into explainable evidence networks for Trust & Safety investigators.

Project Overview

TrustGraph AI is an evidence-grounded marketplace abuse network investigator designed for Trust &
Safety teams. It connects sellers, devices, IP addresses, physical addresses, settlement fingerprints,
listings, product images, reviewers, complaints, and prior cases into an inspectable evidence graph.
The platform combines deterministic relationship analysis, hybrid retrieval, and AI-assisted reasoning. It
never performs automatic enforcement. A human reviewer always makes t

Problem Being Solved

Marketplace abuse often operates as a coordinated network, while day-to-day reviews usually begin with
a single account. A previously removed operator may return using new seller names, email addresses,
and storefronts while quietly reusing devices, IP ranges, addresses, settlement details, listing images,
reviewer accounts, or operational patterns.
Account-by-account review can therefore miss important relationships. Manual correlation is slow, weak
signals may create false positives, and decisions can be difficult to reproduce. TrustGraph converts
isolated review into a connected, evidence-backed investigation.

How TrustGraph AI Works

1. Deterministic evidence layer
Builds exact and derived relationships, transparent risk signals, counter-evidence, and complete
evidence provenance.

2. Retrieval layer
Uses Qdrant dense vector retrieval and BM25 keyword retrieval. Reciprocal Rank Fusion combines the
result sets to surface relevant policies and prior cases.

3. AI investigation layer
A Groq-hosted LLM receives graph evidence, deterministic signals, and retrieved context. It drafts
structured findings that may cite only existing evidence IDs.

4. Evidence critic
Validates evidence references, rejects unsupported claims, and prevents AI output from becoming the
source of truth.

5. Human review
The reviewer may clear the case, request enhanced verification, escalate it, or mark it inconclusive.
Enforcement remains outside the AI workflow.

Technology Stack

Area Technology

API/Workflow

FastAPI, LangGraph
Graph & Storage NetworkX, SQL
RetrievalQdrant dense retrieval + BM25 + Reciprocal R
LLMGroq structured JSON using openai/gpt-oss-20b; mock mode for offline demos
and CI
User Interface Streamlit reviewer co
QualityUnit, integration, security tests, and a 15-case

Project Resources

Resource Link
Githubhttps://github.com/Bindukk18/trustgraph-ai
Setup and demonstration pathhttps://github.com/Bindukk18/trustgraph-ai/blob/main/README.md
Architecture https://github.com/Bindukk18/trustgraph-ai/blob/main/docs/ARCHITECTURE.md
Demo Vediohttps://drive.google.com/file/d/1yRZu8Xy5EhgTR3JWG7SH3UvSiqbKy4DF/view?usp=sharing