AWS Builder Center
GroundTruth: Continuous AI Reality Verification with Amazon SageMaker and Serverless AWS Architecture

GroundTruth: Continuous AI Reality Verification with Amazon SageMaker and Serverless AWS Architecture

GroundTruth is an AI reality verification platform that parses safety SOPs and compliance manuals into visual rules, detects physical hazards using Amazon SageMaker AI multimodal computer vision, and automates closed-loop remediation across 12 native AWS production cloud services.

GroundTruth: Continuous AI Reality Verification with Amazon SageMaker and Serverless AWS Architecture

AWS First Commit 2026 — SHIP IT Track Participant: Pochiraju Kailash Ram Markandeya Sharma (@kailashsharma) Team: KGP_unknown_Coder_404 (Team Code: BFZXQT) GitHub Repository:  GROUNDTRUTH-First-Commit-Hackathon-2026 



1. The Challenge: The Physical "Reality Gap"

Organizations spend millions authoring comprehensive safety protocols, maintenance SOPs, and compliance standards. However, verifying on-the-ground compliance still depends heavily on manual clipboards, unstructured messaging groups, and static spreadsheets.
This creates a dangerous "Reality Gap" — safety standards exist on paper, but physical facilities (warehouses, factories, data centers, hospitals) frequently suffer from unaddressed safety hazards like blocked emergency exits, obstructed electrical switchboards, and missing PPE.
GroundTruth solves this problem by turning static policy documents into continuous, machine-verifiable visual reality checks powered by Amazon SageMaker AI and an event-driven AWS serverless architecture.



2. Platform Workflow: From Policy to Verified Fix

GroundTruth operates across a 5-stage automated pipeline:
  1. Policy & SOP Ingestion: PDF/DOCX safety manuals are ingested and parsed into atomic visual and procedural requirements (e.g., "Emergency exit doorway must remain 100% unobstructed across 1.2m width").
  1. Encrypted Field Evidence Vault: Field inspectors capture photos that are uploaded securely to Amazon S3 via ephemeral, time-bounded Presigned URLs.
  1. Multimodal Visual Reasoning: Amazon SageMaker AI runs spatial and semantic inference on the evidence photo against the policy rule, computing confidence scores, bounding boxes, and natural language explanations.
  1. Duplicate Clustering & Facility Memory: Vector similarity algorithms prevent duplicate tickets and detect recurring incident hotspots at facility locations.
  1. Closed-Loop Remediation & Before/After Engine: Maintenance teams submit post-fix photos. SageMaker AI runs multi-image comparison, presented to human supervisors via an interactive Before/After verification slider, before committing a cryptographically signed SHA-256 record to Amazon DynamoDB.



3. Production AWS Cloud Architecture (12 Native Services)

GroundTruth was built natively for the AWS Cloud using 12 integrated services deployed via Infrastructure-as-Code:
  • AWS Amplify: Edge distribution and hosting for the React + TypeScript frontend application.
  • Amazon API Gateway: RESTful API routing, request validation, and rate-limiting.
  • AWS Lambda (Node.js 20.x): Serverless microservice execution for OCR extraction and business logic.
  • AWS Step Functions: Asynchronous workflow orchestration state machine coordinating upload, AI inference, deduplication, and notification steps.
  • Amazon SageMaker AI: Real-time multimodal PyTorch vision reasoner endpoint for spatial discrepancy detection.
  • Amazon S3: SSE-S3 encrypted object storage vault for policy documents and high-resolution visual evidence.
  • Amazon DynamoDB: Low-latency NoSQL database managing Policies, Findings, Actions, and Immutable Audit Trails across Global Secondary Indexes (GSIs).
  • Amazon Cognito: Role-based access control and JWT authentication for Field Auditors, Safety Officers, and Facility Directors.
  • Amazon EventBridge: Event bus managing decoupled domain events ( GAP_DETECTED ,  REMEDIATION_VERIFIED ).
  • Amazon SQS (FIFO Queue): Ordered inspection processing with an attached Dead Letter Queue (DLQ).
  • Amazon SNS: Immediate push notifications to facility managers when critical safety gaps are identified.
  • Amazon CloudWatch: Centralized telemetry, structured access logs, latency alarms, and operational dashboards.



4. Verification & Quality Engineering

To ensure production reliability, GroundTruth was validated with an automated 1,000-case test suite spanning 7 layers:
  • Section 1: Health, Config & System Endpoints (50 cases) — 100% Pass
  • Section 2: Policy Ingestion & OCR Normalization (150 cases) — 100% Pass
  • Section 3: Field Inspections & S3 Presigned Uploads (200 cases) — 100% Pass
  • Section 4: SageMaker AI Vision Multimodal Reasoning (200 cases) — 100% Pass
  • Section 5: Duplicate Clustering & Vector Facility Memory (150 cases) — 100% Pass
  • Section 6: Remediation Verification & Multi-Image Comparison (150 cases) — 100% Pass
  • Section 7: Human Confirmation & SHA-256 Audit Trail (100 cases) — 100% Pass
Result: 1,000 / 1,000 Automated Test Cases Passed (100.0% Pass Rate).



5. Conclusion

GroundTruth demonstrates how combining Amazon SageMaker AI with serverless AWS primitives transforms compliance from periodic, manual paper audits into proactive, continuous reality verification.
  • Live Demo Video: 2m 29s HD Walkthrough with Neural Audio Narration
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article