SYSTEM READY · 15 AGENTS ONLINE

Criminal networks,
decoded by agents.

Intel-Forge ingests CCTV, FIRs, CDRs and financial logs — fuses them into a live knowledge graph, and answers investigator questions with evidence-backed, cited leads. Every claim carries its receipt.

Drop evidence
Zero-trust citations · HITL quarantine · What-If overlays · court-ready audit trail
📡 CDR burst · +91 98••• 4412
👁 YOLOv8 · person 96.4%
🕸 PageRank · kingpin: R. Kumar
PROTOTYPE · SYNTHETIC DATA
LIVE NETWORK — CASE FIR-2026-118 drag the nodes ✦
Person Phone Account Location Evidence
Multimodal ingestion

The Evidence Desk.

Drop a CCTV still, footage, or a scanned FIR — the Visual, NLP and Validation agents pick it up, extract what matters, and stamp every claim with a deterministic citation. All processing here is simulated for the prototype.

Drop evidence here

or click to browse — images get object detection, documents get OCR + NER, videos get frame scanning.

JPGPNGMP4PDFTXTCSV
Confidence gates enforced. Detections under 60% are flagged, never auto-linked — they route to the HITL quarantine instead.
§
No claim without citation. Every extracted entity ships with an evidence pointer: file, frame, or FIR line number.
// queue empty — awaiting evidence_
LangGraph orchestration

Watch the agent network think.

A query enters the Orchestrator, gets decomposed, and fans out to specialist agents. Each one self-verifies before the Citation agent lets an answer through.

OR

Orchestrator

decomposes & routes
standby
TL

Telecom

CDR sub-graphs
standby
FI

Financial

money-flow tracing
standby
VI

Visual

YOLOv8 · CCTV
standby
NL

Text & NLP

OCR · cross-lingual NER
standby
GD

Graph Sci.

PageRank · Louvain
standby
CT

Citation

zero-trust receipts
standby
Why it matters

Weeks of cross-referencing,
compressed to minutes.

Investigators drown in siloed evidence. Intel-Forge fuses it into one graph and reasons across all of it — with humans in the loop at every threshold.

0M
evidence objects indexed
0
self-verifying agents
0%
citation completeness
<0s
avg time to first lead

GraphRAG reasoning

Multi-hop traversal over Neo4j + Qdrant — answers come from the fused graph, not guesswork.

Zero-trust citations

Every claim carries a Neo4j node ID, FIR line or CCTV timestamp. Unverified claims are blocked before they reach the officer.

Human-in-the-loop

Merges and edges under 85% confidence land in a quarantine queue — an officer approves before anything touches the master graph.

What-If overlays

"Assume this number belongs to Suspect X." Hypotheticals run on ephemeral graph branches — the master DB is never touched.

Cross-lingual NER

Hindi, regional scripts and degraded scans resolve into one unified English entity graph — aliases included.

Immutable audit trail

Every agent action is hash-chained. Tamper with one entry and the whole chain screams — built for court admissibility.