Atlas fuses graph ML with LLM reasoning to explain who is connected to whom – and why it matters. Spot fraud rings, hidden exposure clusters, and systemic risk faster.
Multi‑hop device/identity reuse, mule accounts, common employers, broker hubs.
Natural‑language narratives with evidence and policy mapping.
Portfolio‑level maps of co‑risk clusters and contagion paths.
Agentic OSINT pulls leads & evidence links automatically.
POST /intel/atlas/score
{
"borrower_id": "B890",
"subgraph": {"k_hop": 2, "max_nodes": 200},
"docs": ["stmt_2024_11.pdf", "employment_letter.txt"]
}
200 OK
{
"risk_score": 0.79,
"tier": "high",
"features": [
"Shares device with 4 defaulters",
"Transactions with shell vendor V102",
"Employer linked to 12 risky borrowers"
],
"narrative": "Borrower shares device fingerprint with four prior defaulters and has transacted with vendor V102 flagged in cases C-118 and C-203. Employment cluster overlaps with a high-risk group from 2024Q4.",
"evidence": [
{"type":"device","id":"fp_abc123"},
{"type":"txn","id":"tx_892312"},
{"type":"case","id":"C-118"}
],
"actions": ["HOLD","STEP_UP_KYC","OPEN_CASE"],
"model_version": "atlas_v1.1.0"
}Device→identity: FP clusters, emulator flags
Txn→vendor: Shell vendors & mules
Employer→borrower: Risky employer hubs
IP→borrower: Shared IP/ASN patterns
It computes relational features per case and explains them in natural language – not just “edges”, but the why with evidence and policy mapping.
Subgraph extraction and scoring are optimized. Typical p95 < 400ms for 2‑hop, <200 nodes; larger graphs run async.
GNN/graph ML computes embeddings; LLMs produce factual narratives; agents gather OSINT with guardrails.
Yes – begin with ExplainAI on top of existing rules, then add GraphLLM scoring and AgentOps.
We’ll tailor a walkthrough to your portfolio and data landscape.
© 2025 Deepvue Technologies Pvt. Ltd.