Pilot Rollout Plan
4-Week Pilot Rollout Plan
A structured plan for a successful pilot deployment, data collection, and model improvement.
Objective: Onboard a pilot group of 2-3 project managers and site engineers. Familiarize them with the DPR and Punchlist features.
Data Collection Goals:
- Photos: 20-30 photos per day from 1-2 active projects. Focus on variety: different activities (concreting, plumbing), times of day, and some clear defects.
- Invoices: 10-15 historical invoices. Include a mix of simple and complex formats, handwritten, and scanned copies.
Actions:
- Conduct a 1-hour training session on using the app.
- Users start uploading daily photos and creating DPRs.
- Users upload the first batch of invoices for OCR extraction.
- Begin human-in-the-loop (HITL) verification for all AI outputs.
Human-in-the-Loop (HITL) Verification
Templates and instructions for creating high-quality training data by verifying AI outputs.
Photo Defect Verification
For each photo analyzed by the AI, a human reviewer should fill out a corresponding row in a CSV file like the one below. This data is critical for fine-tuning the vision model.
| photo_id | human_verified_defect_type | human_verified_severity | ai_defect_type | ai_severity | is_correct | notes |
|---|---|---|---|---|---|---|
| img_101.jpg | Improper PPE Usage | high | Improper PPE Usage | high | TRUE | |
| img_102.jpg | Hairline Crack | low | Structural Crack | medium | FALSE | AI misclassified severity and type |
| img_103.jpg | None | n/a | None | n/a | TRUE |
Invoice OCR Verification
For each invoice processed, a reviewer should verify the key extracted fields, especially the line item amounts. Low-confidence items flagged by the AI should be prioritized.
| invoice_id | line_item_desc | human_verified_amount | ai_extracted_amount | is_correct | notes |
|---|---|---|---|---|---|
| INV-001 | Portland Cement | 50000 | 50000 | TRUE | |
| INV-002 | TMT Steel 12mm | 50000 | 5000 | FALSE | AI missed a zero in the amount |