“An AI system that doesn't just generate a root cause—it investigates, challenges, validates, and knows when it doesn't know.”
ReliAI is a mission-critical Multi-Agent Industrial Incident Investigation Harness designed for modern automated manufacturing, assembly lines, and robotics stations (e.g. 6-DOF robotic arm tire-fitting cells).
When an industrial incident occurs (machine stall, sensor anomaly, quality defect, thermal spike, or acoustic grinding), engineers must cross-examine data across multiple disparate systems:
- Real-time PLC & sensor telemetry
- Thermal imaging cameras & acoustic audio recordings
- Maintenance logs & component lifecycle history
- Standard Operating Procedures (SOPs) & Golden Run specifications
Conventional single-prompt AI systems hallucinate plausible-sounding root causes without validating physical evidence. In manufacturing, premature conclusions or false positives lead to catastrophic downtime, equipment damage, and severe safety hazards.
ReliAI decomposes root cause investigation into specialized, evidence-grounded agents with strict adversarial validation:
[ Multimodal Incident ] ➔ [ Triage Agent ] ➔ [ Evidence & RAG Agent ]
│
┌──────────────────────────────┴──────────────────────────────┐
▼ ▼
[ Telemetry & Kinematics Agent ] [ Quality & Fitment Agent ]
│ │
└──────────────────────────────┬──────────────────────────────┘
│
▼
[ Root Cause Generation ]
│
▼
[ Adversarial Critic Agent ]
(Falsification & Contradiction)
│
▼
[ Deterministic Confidence Engine ]
│
▼
[ Human Approval Gate (HITL) ]
- Evidence-Grounded Reasoning: An agent cannot propose a cause without citing specific sensor timestamps or baseline deviations.
- Adversarial Critic Loop: A dedicated Critic Agent actively attempts to disprove the proposed hypothesis by searching for contradictory sensor data.
- Deterministic Confidence: If evidence is missing or contradictory, the system refuses to speculate and outputs:
"INVESTIGATION INCONCLUSIVE — CONFLICTING EVIDENCE DETECTED (HUMAN INSPECTION REQUIRED)". - Human-in-the-Loop Gate: All high-impact mitigations require explicit human authorization before execution.
Explore the detailed system architecture in our dedicated documentation guides:
| Document | Focus Area | Description |
|---|---|---|
| 📱 FRONTEND_UI.md | Mission Control UI | React 19 + Vite SPA, 3D Robotic Arm Digital Twin (Three.js/R3F), Multimodal Inspector (Thermal, FFT Audio, Power), SSE Streaming, and Next.js vs. Vite analysis. |
| 🌐 BACKEND_WEBSITE.md | Web Platform Backend | FastAPI + Uvicorn API gateway, PostgreSQL 16 + pgvector schema, Redis Pub/Sub event bus, human approval audit logs, and edge server resource analysis. |
| 🤖 INVESTIGATION_HARNESS_AI.md | Self-Hosted AI Harness | Multi-agent orchestrator, local Ollama model serving (qwen2.5:7b & deepseek-r1:7b), anti-hallucination critic loop, and hardware performance-to-cost analysis. |
main: System architecture, core design specifications, and central documentation.example-render: Interactive 3D robotic arm digital twin rendering prototype.ai-harness: Active development branch for the Python multi-agent investigation harness, Ollama integration, and validation engine.