Skip to content
English
Contact us
  • There are no suggestions because the search field is empty.

PRD: Vision AI for VPOD Image Validation

1. Executive Summary

One-sentence positioning

Real-time on-device AI validation ensures VPOD images meet quality standards before driver submission, reducing disputes, enabling configurable overrides per customer, and providing operations with comprehensive monitoring and compliance controls.

Problem We’re Solving

Low-quality VPOD images, blurry, dark, poorly framed, with incomplete package or address visibility, force operations teams into manual review cycles post-delivery. This causes customer disputes, support escalations, re-delivery attempts, and operational overhead. Manual auditing of attachments is time-consuming and reactive, often uncovering issues too late to address them efficiently

Proposed Solution (High Level)

Integrate Captur edge AI SDK into CXT Driver app for real-time (live camera) and post-capture VPOD image validation. Enable Operations users to configure which validation rules apply per customer—a single rule for On-Demand work and a single rule for Routed work. Surface overrides and validation results in VDB real-time monitoring, Attachments columns, and aggregated driver metrics. Automatically track overrides with status codes, send configurable notifications (toasts/2hr email summaries), and support optional driver override capability via MOS/Driver permissions. Enable usage-based billing per validation event .

Primary Value Delivered

  • Risk Reduction: Fewer customer disputes and chargebacks from unverifiable or poor-quality PODs.

  • Operational Efficiency: Eliminates post-delivery manual attachment audits; auto-flagging surfaces issues in real time.

  • Revenue Enablement: Usage-based pricing model unlocks new revenue per validation event.

  • Strategic Differentiation: AI-powered VPOD quality control positions CXT as operator-first platform for regulated logistics (pharma, medical).

  • Compliance & Accountability: Auto-generated status codes create auditable trail of override events for regulatory and customer inquiries .

Who This Is For

  • Buyer/Decision Maker: Ops Managers in retail, medical, pharmaceutical, and courier industries requiring reliable POD capture and compliance.

  • Primary Users:

    • Drivers: Capture VPODs with real-time feedback; override if permitted (permission-dependent).

    • Dispatchers: Monitor VPOD quality via VDB grid indicators, override flags, and 30-day override counts.

    • Org Admins: Enable per-customer, configure rules (On-Demand and Routes separately), manage MOS permissions and notification channels.

  • Secondary Users: Customer service teams reviewing proof attachments and handling disputes .

2. Strategic Context & Alignment

2.1 Why Now

Market/Customer Drivers: Rising customer expectations for chain-of-custody documentation, especially in regulated industries (pharma, medical). Manual VPOD reviews create bottlenecks and liability exposure. Captur SDK is proven and production-ready post-pilot; integration timing aligns with broader AI-native operations roadmap.

Technical Enablement: Captur partnership matures with dedicated API keys per customer. CXT Driver app architecture supports edge inference without backend calls. VDB infrastructure handles real-time metric aggregation.

Competitive Imperative: Competitors lack real-time VPOD validation on-device; CXT positioned to lead in last-mile proof quality .

2.2 Strategic Alignment

  • Product Vision: Supports platform evolution to AI-native ops.

  • Brand Voice: Reinforces operator-first positioning—tools that reduce friction for dispatchers and drivers while improving delivery outcomes for enterprise customers.

  • Marketplace Strategy: Complements Marketplace by introducing first native AI module (Vision AI) alongside integrations (GigSafe, etc.).

  • Monetization: Allows usage-based billing alongside subscription tiers;

2.3 Non-Goals (Explicit)

What this project is intentionally not doing

  • Custom rule creation by customers: Rule creation/editing remains backend-only; Phase 2 exploration only.

  • Validation for non-VPOD images: Attachments and Required Driver Input photos are explicitly excluded.

  • Backend-only processing: Edge inference is required; no server-side validation fallback in Phase 1.

3. Problem Statement & Opportunity

3.1 Current State

Drivers capture VPOD images without real-time quality feedback. Operations sees raw image attachments only after delivery completion. Quality issues (blur, darkness) are discovered via manual review, customer complaints, or dispute requests. No visibility into which drivers frequently submit poor images. No scalable way to coach drivers or enforce standards at capture time.

3.2 Impact

  • Business: Disputes/churn from unprovable deliveries; re-delivery costs; regulatory exposure in medical/pharma, where chain-of-custody is critical; support burden increases.

  • User: Dispatchers waste hours auditing; drivers rework stops; drivers often unaware of image failed standards; customers dispute over delivery proof quality

3.3 Opportunity

By validating VPOD images at capture time, CXT can enforce quality gates on the driver device before submission. Real-time feedback enables self-correction. Operations gains instant visibility via VDB monitoring, reducing manual audits.

4. User Experience: End-to-End Journey

This section describes how users discover, enable, and live with the feature over time.

4.1 Feature Discovery

  • Org Admins (Operations): "Enable real-time AI validation for VPOD images" toggle appears in Customer > Shipments tab, Validation Rule selectors (On-Demand Validation Rule, Route Stops Validation Rule).

  • Dispatchers (Operations VDB): After enablement, new columns appear in the Shipments grid (VPOD Override Yes/No/Blank with count); the Route Manifest gains per-stop override indicators (flag icons); the Drivers grid displays "VPOD Overrides (30d)" count column.

  • Drivers (CXT Driver App): VPOD capture camera preview shows real-time validation feedback (red/green outlines, status text); Guidelines pop-up on first VPOD capture: "Visual PODs must meet guidelines...".

4.2 Activation Access Control

Who Can Enable:

  • Org Admins: Customer-level toggle "Enable real-time AI validation" + On-Demand/Route Stops rule selectors (e.g., Standard Quality Check, or No validation).​

  • Dispatchers/Fleet Managers: MOS/Driver permissions (Allow VPOD Override, in-app override toasts notifications/ 2-hour summary emails).​

  • Drivers: No explicit activation; inherits via MOS sync.

How It Is Enabled:

  1. Admin enables per-customer -> validation rules (On-Demand and Routes) auto-sync to the Driver app on next login.

  2. Dispatchers configure MOS/Driver override permissions and notification channels.

  3. On next VPOD capture, driver app loads customer rule and enforces validation workflow.

  4. All validation results, overrides, and metadata sync to backend on stop completion .

Key Decisions: Separate On-Demand vs. Routes rules; Single rules for On-Demand and Routes (simpler for Ops, easier to manage); Drivers' override capability OFF by default.

4.3 Post-Activation Experience

  • Drivers (CXT Driver App):

    • VPOD Workflow: Trigger on Camera VPOD capture→ Guidelines popup → Live camera (red/green outlines: "Image too blurry/dark" vs. "Good") → Capture → Post-result:

      • Pass: Green, "Image quality is good" → Retake/Use Photo.

      • Fail: Red, reason -> Retake/Use Photo (if permitted).

      • Error: "Validation unavailable" → Retake/Use Photo (always).

    • Override: Confirm dialog ("Image doesn't meet the quality standards. Do you want to proceed?") -> Metadata recorded (Override Used = Yes), Status code ('VPOD Image Override Validation'/ 'VPOD Image Validation Errored') added.​

    • Offline: Full workflow; syncs metadata later.​

  • Org Admins/Dispatchers (Operations):

    • Customer Config: Shipments tab displays rule selectors with descriptions (e.g., "Not too dark/blurry"); save propagates to all orders/routes; takes effect on next app sync.​

    • VDB Real-Time:

      • Shipments Grid: New column ‘VPOD Override’ (Yes / No / Blank), color-coded, sortable.​

      • Route Manifest View: New column ‘VPOD Override’, per-stop flags.​

      • Drivers Grid: New column: VPOD Overrides (30d) - rolling 30-day count of override events per driver.

    • Order/Stop Details: Attachments section + New columns (Type, AI Validated, Validation Result, Override Used); Summary strip ("4 validated: 3 Pass, 1 Override"); Override rows highlighted.​

    • Notifications: Toasts ("Override Driver 100, Stop 12") fleet-filtered; 2hr emails (Summary counts, Detailed table: Driver ID, Driver Name, Order ID / Route Stop, Customer, Event Time, Validation Outcome, Override Type, Image file/description, Validation Rule, Validation Response.).​

    • Status Codes: System auto-adds status code ’VPOD AI Override’ on submit; removes on image delete (pre-completion).​

  • Day-to-Day Changes:

    • Drivers: Friction on poor images → better proofs with real-time feedback.

    • Ops: Proactive spotting of high-override drivers, Real-time notifications​, Reduced manual attachment review while VDB flags surface outliers, Attachments grid provides one-stop proof, quality audit, Status codes create audit trail for compliance/dispute resolution

  • Failure Edge States:

    • SDK Error: Use Photo always available and Driver proceeds w/ "VPOD Image Validation Errored." code; System records: Validation Result = Error, Override Used = No, Ops sees Error in Attachments.​

    • High Overrides: 30d override count visible in Drivers grid, Notifications help spot consistent patterns.​

    • Validation Rule Removal / Configuration Unavailable: If assigned rule removed from tenant, system falls back to No validation.

  • Confidence Signals: VDB summary strips showing validation counts & outcomes, Override badges & row highlighting in Attachments grid, Metadata visibility (AI Validated, Result, Override Used columns), Status codes with audit trail, Driver grid 30d override metric as coaching trigger .

5. Scope & Phasing

Phase 1 — MVP / Initial Release

  • On-device Captur SDK integration (live + post-capture validation).

  • Real-time AI feedback (Retail validation model - blur/darkness checks, no package/address visibility in Phase 1).

  • Customer-level rule configuration (single On-Demand rule, single Routes rule).

  • MOS/Driver-level override permissions.

  • Configurable notifications (toasts, 2-hour emails).

  • Attachments columns (Type, AI Validated, Result, Override Used).

  • VDB monitoring (Shipments, Route Manifest, Drivers 30d grid).

  • Auto-status codes (add/remove on delete pre-completion).

  • Usage tracking per validation event (billing ready).

  • Exclusions:

    • No custom rule UI (backend-only rule management).

    • No validation for Attachments or Required Driver Input photos.

    • No in-app model request workflow in Operations.

    • No analytics dashboard.

Phase 2 — Future Enhancements

  • AI model management, new models per industry, and customer requirements

  • In-Ops new model requests.

  • Analytics dashboards.

  • Validation beyond VPOD

6. Functional Requirements (High Level)

Core Behaviors:

  1. On-Device Validation: Captur SDK runs inference on driver device during VPOD capture (live preview + post-capture). No backend validation call required .

  2. Configuration Scoping: Customer rules (On-Demand, Routes) applied per work type determine which model invoked .

  3. Override Gating: MOS/Driver permission Allow VPOD Validation Override gates whether driver sees Use Photo button on Fail .

  4. Metadata Persistence: Validation result, override flag, rule name, response text persisted with image attachment .

  5. Status Code Automation: System auto-adds VPOD AI Override code on override; auto-removes on image delete (pre-completion) .

  6. Aggregation: 30-day override counts per driver, queryable for VDB grid .

  7. Notification Routing: Toast/email sent if driver Enable Override Notifications is ON; respects fleet filters .

  8. Offline Capability: Full validation workflow cached locally; syncs metadata on reconnect .

Required Automations:

  • Rule sync to driver app on login.

  • Validation result capture & metadata storage.

  • Status code auto-generation (system-flagged).

  • 2-hour email job (scheduled).

  • Toast delivery (real-time to VDB users).

  • 30-day rolling count aggregation.

Gating/Enforcement Logic:

  • If customer AI validation disabled: no validation invoked, image accepted.

  • If rule = No validation: no validation invoked, image accepted.

  • If rule selected + validation fails + override not allowed: Retake button only; no Use Photo.

  • If override allowed: Use Photo button available; confirmation dialog on selection.

  • If SDK error: Use Photo always available (graceful fallback).

7. System Architecture & Technical Considerations

7.1 Systems Involved

  • CXT Driver: Captur SDK (edge inference).

  • Operations: Customer config UI (Customer > Shipments tab). MOS/Driver permission UI. VDB grids & Order/Stops detail (Attachments) views.

  • Backend: Stores image metadata (AI Validated, Result, Override Used), Auto-adds/removes VPOD AI Override status code jobs, notification queues (triggers toast & 2-hour email jobs), Aggregates 30-day override counts.

  • Third-party (Captur): API keys per customer; model inference SLA; error handling.

7.2 Data Ownership & Flow

  • Read: Driver app reads customer rules (On-Demand, Routes) from backend on login → loads Captur models locally; VDB reads Attachments metadata, Status Codes, 30-day override aggregates on demand.

  • Write: Validation result, Override Used flag, metadata → syncs to Attachments on stop completion; Status codes (auto-generated), notifications (queued events)

  • Async: Notification jobs (2hr emails).​

7.3 Key Technical Decisions

  • Usage metering per customer/model.

8. Data Model & Schema Strategy

Standardized Fields (Attachments):

  • Type (Camera VPOD/ Image/ Others).

  • AI Validated (Yes/No).

  • Validation Result (Pass/Fail/Error/ Not Validated).

  • Override Used (Yes/No/NA).

Extensibility

  • Rule name/response text update without schema change (backend-managed).

  • New validation result types added as enum values in future.

  • Custom metadata fields can be added in metadata JSON blob .

9. Monetization, Packaging & Enablement

9.1 Pricing & Packaging

  • Usage-based: $X per validation event (tiered volume discounts).

  • Charge Trigger: Every time AI validation runs

  • Trial: 1k free events

9.2 Sales & Support Enablement

Sales Positioning:

  • Highlight risk reduction (fewer disputes) and efficiency gains (no manual audits).

  • Position for regulated verticals (medical, pharma) as compliance advantage.

  • Reference Marketplace integration (co-sell with other modules).

  • ROI calculator: "At 10,000 deliveries/month with manual audit cost $2/order, Vision AI saves $20k/month in labor + reduces chargeback risk."

Support:

  • CXT Support owns Ops config & notification troubleshooting.

  • Captur Support owns SDK errors, model accuracy issues (escalation path defined).

10. Success Metrics & Guardrails

Primary Success Metrics

  • Adoption: 20 customers enabled (180 days post-launch).

  • Quality: Override rate <10% (Low rate = high-quality driver images; high rate = rule tuning or training opportunity).

  • Revenue: $X MRR from Vision AI events (12 months).

Operational Guardrails

  • SDK error <1%;

  • Notification delivery 99%

11. Risks, Tradeoffs & Open Questions

Open Product Decisions

  • Exact usage metric (per event vs. per customer)?

    • : We should be tracking the number of images that are audited (per customer) for the usage metric. This is the unit for which we are financially accountable to Azure.

    • In future phases, if we use the Custom Vision API, there may be retraining costs to account for

Technical Risks

  • Captur SDK performance regression: Likelihood - Low, Impact - High, Mitigation - Monitor SDK latency; fallback to error state if > 5s.

  • Validation metadata bloat: Likelihood - Medium, Impact - Medium, Mitigation - Partition Attachments table by date; archival policy.

  • Override notification spam: Likelihood - Medium, Impact - Medium (user disengagement), Mitigation - Fleet filters + opt-in channels.

12. Delivery Plan & Epics

Primary Epic(s)

Target Milestones

  • Phase 0 (Internal): Internal testing and Pilot with Dropoff with limited model capabilities - Month 1 (by Jan 2026)

  • Phase 1 (MVP): Retail industry validation model for demo at FMF conference and for Dropoff’s production usage - Month 2 (by Feb 2026)

  • Phase 2 (Enhancements): Customer demos and marketing, Full public launch, Working on enhancements - Month 3 & 4 (by Apr 2026)

13. Appendix & Links

PRD Quality Bar (Checklist)

Before marking this PRD “Ready”:

  • Clear user journey from discovery → mature state
  • Explicit non-goals documented
  • System boundaries defined
  • Activation and permissions clearly stated
  • Success metrics agreed upon