User Guide

The LDS AI Portal provides a unified AI Services interface for testing multiple AI-powered vision services. Images are routed through the LDS AI API to the appropriate engine — Google Vision for VIN OCR extraction, or AWS SageMaker AI for jumped chock detection.

Use the service tree sidebar to select a service category and choose a specific service to begin testing. Active services include VIN Intelligence → Single OCR, VIN Intelligence → Provider Benchmark, Damage Detection → Jumped Chocks 60, Damage Detection → Damage 10-34, and Damage Detection → Multi-Damage 03-07. Other services (VIN Cropper, Bulk OCR, RailOps, Barcode, Container OCR, Railcar OCR) are planned or in development.

1. Navigate to AI Services from the top navigation bar, then select VIN Intelligence → Single OCR in the service tree.
2. Fill in the Inspection ID (required), Customer, Inspection Type, and vehicle details.
3. Click the upload area and select a vehicle image (JPEG, PNG, or WebP).
4. Wait for the image to upload — a green checkmark confirms success.
5. Assess image quality and VIN visibility in the Image Quality Review section.
6. Click Run OCR to process the image through the LDS AI API.
7. Review the extracted VIN and OCR results in the right panel.
8. Enter an Inspection ID and click Save Result to store the record.

1. Navigate to AI Services, then select Damage Detection → Jumped Chocks 60 in the service tree.
2. Click the upload area and select a vehicle image (JPEG, PNG, or WebP, max 10 MB).
3. The image uploads directly to secure private S3 storage via a presigned URL — wait for the green confirmation.
4. Fill in optional Inspection ID, Inspection Type, and VIN (17-character, omit if unknown).
5. Click Run Detection to process the image through the LDS AI API (AWS SageMaker AI engine).
6. Review the detection result, vehicle count, confidence, and annotated images in the right panel.
7. Click Save Result to store the record. Only completed detections can be saved.

Detection results show one of the following statuses:

• Detected — A jumped chock was detected in the image.
• Not Detected — No jumped chock was found.
• Needs Review — The model flagged the result for manual review (highest priority).
• Failed — Processing failed. Check image quality and retry.

Multi-Damage Detection (03-07) identifies Chip and Gouge damage across 48 supported vehicle areas. Each detected damage item is classified with a severity level of 1–5 and an AIAG damage code for standardized reporting.

1. Navigate to AI Services, then select Damage Detection → Multi-Damage 03-07 in the service tree.
2. Click the upload area and select a vehicle image (JPEG, PNG, or WebP, max 10 MB). The image uploads to secure private storage.
3. Fill in optional Inspection ID, Inspection Type, and VIN (17-character, omit if unknown).
4. Click Run Detection to submit the job. The status shows Processing while the AI model works.
5. Review the detection outcome, damage count, evaluation summary, and annotated images in the results panel.
6. Click Save Result to store the record. Only completed detections can be saved.

Possible outcomes:

• Damage Detected — One or more damage items were found, each with a part assignment, damage type, severity, and AIAG code.
• No Damage — The model found no damage in the image.
• Needs Review — The model flagged the result for manual verification. Always inspect the image manually before accepting.

Result images: The results panel displays three annotated images:

• Original — The uploaded image as submitted, with no overlays.
• Parts Detection — The image with detected vehicle parts outlined and labeled.
• Damage Detection — The image with bounding boxes around each detected damage item.

Processing time: When the AI model is starting from zero (cold start), it may take approximately 8–11 minutes to complete. The status remains Processing during this time — this is normal and not a failure.

Do not resubmit or re-upload the image while the job is Processing. The portal continues polling the same job automatically until it reaches a completed or failed state.

About confidence percentages: Confidence values shown for part and damage detection are AI confidence indicators, not guarantees of accuracy. Results marked Needs Review should always be manually verified against the image before being accepted.

The Provider Benchmark compares VIN OCR extraction across multiple providers side-by-side against a ground-truth VIN.

1. Navigate to AI Services, then select VIN Intelligence → Provider Benchmark.
2. Upload a VIN image — it uploads to private S3 storage.
3. Select providers to include in the benchmark (only connected providers can be selected).
4. Click Run Benchmark — each selected provider processes the image concurrently.
5. Review Provider Results cards showing extracted VIN, confidence, and processing time.
6. Check the Comparison Summary table for a side-by-side match against ground truth.
7. Enter the Confirmed VIN (Ground Truth), reviewer name, confirmation date, and image quality.
8. Click Save Benchmark Result to store the comparison.

The Aggregate Benchmark Summary below the workflow shows cumulative accuracy metrics across all saved benchmarks. The Saved Benchmark Results section lets you browse and delete past benchmark records.

• JPEG (.jpg, .jpeg)
• PNG (.png)
• WebP (.webp)

For best results, use clear, well-lit images where the VIN sticker or plate is fully visible and in focus. Avoid glare, shadows, or partial obstructions.

After running OCR, the result will show one of the following statuses:

Confidence scores range from 0–100% and indicate the OCR engine's certainty in the extracted text. Higher scores generally mean more reliable results.

Confidence may display as "—" (em dash) when the API does not return a confidence value. This does not indicate a failure — it means the provider did not supply a score for this request.

Note: A high confidence score does not guarantee a valid VIN. Always verify the extracted VIN against the image.

1. After OCR completes, compare the Extracted VIN with the VIN visible in the image.
2. If the extracted VIN is correct, proceed to save without changes.
3. If incorrect, enter the correct VIN in the Corrected VIN field (uppercase, 17 characters).
4. The corrected VIN will be saved alongside the original OCR output.
5. Saved results can be edited from the Results page.

The Results page displays all saved test records — both VIN OCR and Jumped Chock results — in a single searchable, filterable table.

• Use the Service filter to show All Services, VIN OCR only, or Jumped Chocks only.
• Use the search bar to find records by Inspection ID or extracted VIN.
• Filter by Inspection Type, Result, Review Status, Tester, Source, or Date.
• The Result filter is grouped by service: VIN OCR (Pass, Fail, VIN Not Found, Corrected), Jumped Chocks (Detected, Not Detected), and Both Services (Needs Review).
• Click a row to expand and view full details, including the image, raw OCR output, or chock detection metrics.
• If an Inspection ID has multiple tests, a badge shows the count — this does not block submission.
• VIN review fields (corrected VIN, test status, notes) and Chock review fields (decision, notes) can be edited from the expanded row by authorized users.

The Export page allows you to download test results as CSV files. It has three tabs:

• VIN OCR — Inspection ID, Customer, Provider, VIN, Confidence, Status, and more.
• Jumped Chocks — Inspection ID, VIN, Detection Result, Confidence, Vehicle Count, and more.
• Benchmark — Per-provider comparison: Extracted VIN, Exact Match, Confidence, Processing Time, and more.

For each tab, click Load Data to fetch records, review the summary count, then click Download CSV to export. Use the refresh button to reload the latest data.