AndonAI

System Architecture

AndonAI multi-agent pipeline architecture and AMD GPU acceleration design

Agent Pipeline Architecture

Input Layer

Multimodal evidence ingestion

Camera feedsIoT sensorsLog filesDocumentsManual reports

Watcher Agent

Real-time anomaly detection

Object detectionThreshold monitoringPattern matching

Interpreter Agent

Contextual analysis

Scene understandingRisk assessmentHazard classification

SOP Agent

Procedure matching

SOP database lookupCompliance checkRegulatory match

Root Cause Agent

Causal analysis

Historical correlationPattern analysisContributing factors

Dispatcher Agent

Severity & assignment

Severity scoringTeam assignmentEscalation routing

Report Agent

Audit-ready documentation

Report generationEvidence packagingCompliance formatting

Dashboard

Command board visualization

Live Andon boardIncident trackingAnalytics & trends

AMD GPU Acceleration

Designed for AMD Developer Cloud & AMD Instinct MI300X

AMD Instinct MI300X

192GB HBM3 memory, up to 1,307 TFLOPS FP16 performance for multimodal AI inference at scale.

ROCm Open Platform

Open-source GPU computing platform enabling portable AI workloads with PyTorch and ONNX Runtime.

AMD Developer Cloud

Cloud-based access to MI300X accelerators for development, testing, and production deployment.

Batch Processing

Parallel video frame analysis, embedding generation, and multi-agent workload distribution.

Performance Benchmark

CPU vs AMD Instinct MI300X GPU comparison

Demo benchmark placeholder — Replace with real values via environment variables or the AMD inference service.

CPU Baseline
AMD MI300X GPU
13.0x

Single Frame Analysis

2340ms → 180ms

14.1x

Multimodal Inference

4500ms → 320ms

14.1x

Batch Video Processing (60 frames)

12000ms → 850ms

18.9x

Embedding Generation (1K docs)

1800ms → 95ms

Hardware: AMD Instinct MI300X (Simulated)Simulated benchmark values

AMD GPU Workload Distribution

Batch Video Frame Analysis

Process 60+ camera frames in parallel using ROCm-optimized ONNX Runtime on MI300X 192GB HBM3 memory.

Multimodal Inference

Run vision-language models for scene understanding, object detection, and hazard classification at scale.

Embedding Generation

Generate text and image embeddings for SOP matching and similarity search across the knowledge base.

Agent Workload Parallelization

Distribute independent agent tasks across GPU compute units for sub-second analysis pipeline completion.

Deployment Architecture

AMD Developer Cloud

Cloud-based MI300X instances for development, benchmarking, and production AI inference workloads.

ROCm Stack

Open-source GPU platform with PyTorch, ONNX Runtime, and vLLM for serving multimodal models.

Inference Service

FastAPI-based Python service with /analyze-frame, /batch-analyze, and /benchmark endpoints for GPU inference.

Frontend Application

Next.js 14+ dashboard consuming inference results via REST API. Deployable on any cloud or edge infrastructure.

For Judges: Benchmark values shown are simulated placeholders. Real benchmarks can be supplied via environment variables (BENCHMARK_FRAME_CPU, BENCHMARK_FRAME_GPU, etc.) or by connecting the AMD inference service running on AMD Developer Cloud with actual MI300X hardware.