Product Spec - Edge AI Vision Validation for VPOD Images

Executive Summary

CXT is introducing real-time AI-powered validation of camera-based Visual Proof of Delivery (VPOD) images using Captur AI edge processing. This feature enables customers to automatically assess proof-of-delivery photo quality, allow controlled driver overrides when permitted, and provide Operations teams with visibility into validation results and override patterns via Operations and Visual Dispatch Board (VDB).

The validation logic is fully defined inside each AI model; customers can choose which model to use per end-customer but cannot toggle individual rules within a model. Phase 1 will deliver core functionality in January 2026, followed by a pilot with 2–3 high-value customers in Q2.

1. Business Problem & Opportunity

Today, VPOD images are not automatically validated for quality, leading to disputes, manual review effort, and inconsistent standards across customers. The new AI validation capability targets industries like retail and medical where high-quality proof of delivery is critical and dispute volume is high.

The opportunity is to reduce POD disputes and Operations review time while creating a differentiated, AI-enhanced last-mile experience for CXT customers.

2. Proposed Solution – Phase 1

High-Level Flow

1. Configure

        ↓

Customer specifies AI validation rules per end-customer (Captur-trained models)  

        ↓

1. Validate

        ↓

Driver captures VPOD in CXT Driver → Real-time on-device AI validation 

        ↓

1. Result

        ↓

Pass → Accept | Fail → Retry or Override (if permitted) 

        ↓

1. Visibility

        ↓

Operations see validation status + override indicators in grids + receive alerts

A. Customer-Level Configuration

• Enable/Disable per end-customer:

  • Toggle AI validation for VPOD images at the customer level.

  • If disabled, VPOD behaves as it does today with no AI validation.

• Model Selection (use vs. not use):

  • Customer selects one active model from those provisioned in their environment (e.g., “Retail POD”, “Medical POD”).

  • Each model has predefined validation criteria baked in; customers cannot enable/disable individual rules within the model.

  • The choice is binary: use the model as-is or do not use it for that end-customer.

• Change in rules / criteria:

  • If the business needs different validation criteria, they must raise a request with CXT to provision a new model/API key that reflects the revised validation rules.

  • CXT will work with Captur to train and deploy the appropriate model for that tenant.

• Override Permission (MOS/Driver level):

  • Configuration determines whether drivers are allowed to override failed validations.

  • This is controlled via MOS/Driver rules tied to the customer’s configuration.

B. Real-time Validation on CXT Driver (Designs)

• Trigger:

  • When a driver captures a VPOD image using the CXT Driver app.

• Processing:

  • Captur’s AI model runs on-device validation in real time.

  • No network call is required for the validation step itself.

• Outcomes & UX:

  • Pass: Image accepted; driver proceeds.

  • Fail: Driver sees a message indicating a quality issue and is prompted to either retake or override if overrides are allowed.

  • Error: Technical failure in validation, with a prompt to proceed without validation if necessary.

• Metadata:

  • Each image stores validation result (Pass/Fail/Error/Not Validated) and an override flag where relevant.

C. Operations Visibility

• Attachments (Orders/Stops):

  • New columns to display whether an image was AI-validated, the result, and whether an override was used.

  • Overridden images are visually highlighted so Operations can easily identify risky proof.

• VDB Shipments Grid (On Demand):

  • A “VPOD Override” indicator column flags shipments where drivers submitted VPOD images after failing validation and choosing override.

• VDB Route Manifest:

  • A per-stop “VPOD Override” indicator provides route-level visibility for dispatchers.

D. Notifications & Alerts

• Configuration (MOS/Driver level):

  • Ability to enable or disable override notifications.

  • Option to send 2‑hour summary emails and/or show in-app toasts in VDB when overrides occur.

• Notification Content:

  • Emails summarize recent overrides with key details such as driver, customer, order/stop, time, and validation context.

  • Toasts appear in VDB and show concise info tied to the relevant driver’s fleet.

3. AI Model & API Key Management

Phase 1 – Backend Provisioning

• Model Training & Ownership:

  • Captur trains the AI models based on customer-specific validation needs.

  • CXT manages deployment of these models and corresponding API keys into each tenant’s environment.

• Per-tenant Isolation:

  • Models and API keys are provisioned per tenant/environment, with no sharing across customers, supporting usage tracking and potential billing alignment.

• Naming & Selection:

  • CXT assigns friendly names to models during backend provisioning to reflect the customer’s business context (e.g., “Retail POD”).

  • Operations users can select one active model per configured customer but cannot modify the validation criteria inside that model.

Phase 2 – Future Model Management UI

• Planned capabilities (not part of Phase 1 delivery date):

  • UI to view models available in a customer’s environment.

  • Ability to rename models, enable/disable models, and see a description of the validation criteria associated with each model.

  • Ability to submit a request for a new model/API key with required validation criteria, which CXT will fulfill with Captur and then provision per tenant.

4. User Personas & Workflows

CXT Customer – Account Admin

• Decides which end-customers should use AI VPOD validation and which model fits their industry.

• Requests new models from CXT when new or revised validation criteria are needed.

Operations User – Dispatcher / Ops Manager

• Enables or disables AI validation for specific end-customers.

• Chooses the model to apply (without altering internal rules).

• Configures override behavior and notifications.

• Monitors validation and overrides in Attachments and VDB.

CXT Driver – Driver App User

• Captures VPOD photos and receives immediate pass/fail/error feedback.

• Retakes or overrides failed images depending on allowed configuration.

5. Feature Breakdown – (Epic)

6. Success Metrics & KPIs

Metrics are focused on meaningful adoption and quality outcomes rather than complex dashboards in Phase 1.

• VPOD images passing AI validation: Target >85% among validated images.

• Override rate: Target <10% of validated images using override.

• POD dispute reduction for pilot customers: Aim for 25–30% reduction.

• Operations time spent reviewing attachments: Target a reduction compared to pre‑launch baselines.

• Feature adoption target: 20 customers enabling AI validation within 6 months of general availability.

7. What’s Included in Phase 1

• Real-time AI validation for VPOD images in CXT Driver using Captur on-device models.

• Per-customer configuration to enable/disable validation and select a model on a use vs. not use basis.

• Override workflows controlled via MOS/Driver configuration.

• Operations visibility into validation and overrides in Attachments, VDB Shipments, and VDB Route Manifest.

• Configurable notifications (2-hour email summaries and in-app toasts) on overrides.

• Backend-managed model and API key provisioning per tenant/environment.

8. What’s Out of Scope (Phase 1)

For clarity, Phase 1 does not include:

• Self-service UI for model management (view/rename/enable/disable/request models).

• Any real-time or historical dashboard for trends, pass/fail analytics, or override analytics.

• Inline rule-level configuration inside a model (e.g., toggling “Package Visibility” or other internal rules).

• Validation for non-VPOD images.

9. Rollout & Timeline

• Phase 1 – Core Functionality:

  • Targeted for January 2026 delivery of core VPOD AI validation capabilities.

• Pilot (Q2):

  • Pilot with 2–3 high‑value customers in Q2, focusing on Retail and Medical use cases where proof quality is critical.

• Future Phases:

  • Self-service model management UI and any advanced analytics or dashboards can be considered for subsequent phases after pilot feedback.

10. Success Definition

This feature is considered successful when:

• Adoption: 20 customers enable AI validation within 6 months of general availability.

• Quality: A high percentage of VPOD images pass validation while maintaining a low override rate.

• Business Impact for Pilot Customers: Noticeable reduction in POD disputes and manual review time for Operations teams.

• Stability: No regressions in existing VPOD workflows for customers not using AI validation.

11. Appendices Available Upon Request

• Detailed story PDFs for:

  • Operations configuration.

  • CXT Driver AI validation flow.

  • Attachments and VDB indicators.

  • Notifications.

• Supporting documentation from Captur (SDK/technical integration notes).