Active Project — Meta

D2C Returns Discrepancy Management

Transform manual discrepancy management for Meta product returns into an AI-powered automated process. Covers D2C unit-level returns, B2B bulk shipment receiving, and dock log / trailer-level quantity verification.

D2C ReturnsB2B ReceivingDock VerificationVision ModelsAuto-Adjudication
Current State

D2C Receiving Process Flow

The current process from dock to disposition — each step is largely manual and sequential, with non-value-add dwell time that the AI system will eliminate or minimize.

STEP 1

FedEx truck arrives at dock

STEP 2

FedEx scans package tracking ID

STEP 3

Package placed on flex conveyor

STEP 4

Operator scans tracking ID + RMA into Tracking Tool

STEP 5

Tracking Tool requests FedEx data (reference #, weight, dimensions)

STEP 6

Operator performs Dock Log — bulk upload to Plus

STEP 7

Package moved to receiving line

STEP 8

Receive Operator performs system receiving

STEP 9

Exception → Discrepancy ticket in Plus → Problem Log

Current State Pain Points

Eight critical issues driving the transformation — each one maps to specific future-state automation.

Plus/Jira integration gap

Impact: Manual copy/paste; data entry errors; duplicated effort

Root Cause: No API or automated sync between Plus and Jira

High CSR turnover rate

Impact: Institutional knowledge lost; inconsistent execution

Root Cause: Complex manual process dependent on tribal knowledge

Cases not closed in Plus

Impact: Aging inventory; inaccurate open case counts

Root Cause: Closure requires manual CSR action; no auto follow-through

Security risk: un-wiped units in outside storage

Impact: Data security exposure; compliance risk

Root Cause: Problem Log overflow; units stored outside secure area

Opaque Meta decision rules

Impact: Cannot predict or auto-resolve common cases

Root Cause: Meta's adjudication criteria not shared with RXT

No response pattern analysis

Impact: Every ticket requires manual back-and-forth

Root Cause: No trailing-period analytics on Meta responses

Manual dock quantity verification

Impact: Miscounts; delayed discrepancy detection

Root Cause: No camera or vision-based validation at dock

B2B labels scanned manually

Impact: Delayed discrepancy identification

Root Cause: No vision model to read box labels or count

Future State Architecture

AI-Powered System Components

Eight interconnected components that transform the discrepancy management process from manual to intelligent automation.

Phase 1

Automated Mismatch Detection

System instantly creates match/mismatch events when Serial #, RMA, and Airway Bill are scanned. On mismatch, a Jira ticket is auto-created with full discrepancy context (SN, RMA, model, mismatch type, product photos).

Phase 1

Real-Time Plus–Jira Sync

Bidirectional API integration eliminates manual copy/paste. Jira ticket data is simultaneously transmitted to Plus in real time for claims tracking and insertion into the 2x/day Meta submission.

Phase 3

Auto-Adjudication Engine

Trailing 30/60/90-day analysis of Meta response patterns enables automatic disposition for high-confidence cases. 'New RMA Assigned' and 'Scrap' dispositions are handled automatically when confidence is high.

Phase 2–3

Vision Model — Scrap Classification

AI vision model inspects unit condition. If hazardous goods criteria are met and no Legal Hold applies, scrap is auto-initiated. Historical scrap records train and validate classification accuracy.

Phase 2

B2B Vision-Enhanced Receiving

Cameras scan box labels at point of receipt. System extracts model, quantity, and part number data and compares against expected RMA manifest. Discrepancies reported to Meta in real time.

Phase 2

Dock Camera & Trailer Verification

Hard-mounted cameras in each dock bay with system-driven triggers capture images and record metadata. Vision model auto-counts and cross-references manifest data. Damaged pallets auto-trigger Meta notification.

Phase 3

Agent Auto-Close & Routing

AI agent closes records in both Jira and Plus upon disposition, removing serial numbers from aging inventory reports. Material handlers receive automated pull tickets with rack location and routing instructions.

Phase 1

Secure Unit Routing Protocol

Automated routing ensures units stay in secure area until data wipe is confirmed. Eliminates the current security gap of un-wiped units in outside storage.

Meta Engagement

Meta Feedback Architecture

Structured touchpoints ensuring Meta has full visibility and opportunity to provide feedback at every stage of the transformation.

Weekly

Weekly Progress Reviews

Scheduled review sessions where Reconext presents implementation progress, auto-adjudication accuracy metrics, and upcoming milestones. Meta provides feedback on priorities and rule refinements.

Continuous

Shadow Mode Validation

During Phase 3, auto-adjudication runs in parallel with manual process. Meta reviews AI decisions vs. their actual responses to validate accuracy before going live.

Phase 1 (Critical)

Adjudication Rule Sharing

Meta shares decision rules for common dispositions (New RMA, Scrap, Return to Customer). These rules are encoded into the auto-adjudication engine with Meta's approval.

Phase 1 (Blocker)

Legal Hold Criteria Definition

Joint working session to define Legal Hold vs. No Legal Hold determination criteria — the key gate for automated scrap disposition.

Phase 4

Self-Service Data Interface

Meta receives a generative AI interface to query discrepancy data, resolution patterns, and aging metrics directly — reducing back-and-forth requests.

Monthly

Accuracy & KPI Reporting

Automated reports showing auto-adjudication accuracy, aging reduction, cycle time improvement, and labor savings — shared with Meta on a defined cadence.

Implementation Roadmap

Phase 1

Foundation & Quick Wins

Build Plus–Jira API integration for bidirectional real-time sync
Automate discrepancy case creation and Jira ticket generation on mismatch events
Automate case closure in Plus when Jira ticket is resolved (address aging backlog)
Build trailing 30/60/90-day response analytics dashboard
Establish secure unit routing protocol (prevent un-wiped units from leaving Problem Log area)
Define Legal Hold vs. No Legal Hold criteria with Meta; request Meta's adjudication rules
Phase 2

Vision Model Deployment

Deploy vision-model-enhanced cameras at B2B receiving stations
Install hard-mounted cameras in each dock bay with system-driven triggers
Build manifest scanning capability (visual airway bill extraction)
Train scrap classification model on historical data
Pilot vision-based B2B discrepancy detection (model/mix/qty)
Deploy conveyor-based package counting and label capture system
Phase 3

Auto-Adjudication Engine

Deploy auto-adjudication for 'New RMA Assigned' dispositions
Deploy auto-adjudication for 'Scrap' (with Legal Hold gate)
Build agent logic to auto-close records and clear aging serial numbers
Automated pull ticket generation for material handlers
Measure and validate auto-adjudication accuracy vs. manual outcomes
Phase 4

Optimization & Scale

Expand auto-adjudication to additional response types based on pattern clustering
Continuous learning loop: retrain models on new response data
Extend vision model to D2C unit condition assessment
Full dock-to-disposition automation with exception-only manual intervention
Reporting: aging reduction, cycle time, labor savings, accuracy KPIs
Team Assignments

Exactly What Each Person Does

Every Dream Team member has explicit, actionable responsibilities for this project. No ambiguity — here is exactly what each role owns.

AI Model Coach / Process Optimization Lead

Operations / Engineering

Specific Deliverables & Actions

Validate auto-adjudication accuracy by comparing AI decisions against actual Meta responses
Define and refine the confidence thresholds for auto-adjudication (when is 'high confidence' high enough?)
Interpret vision model scrap classification outputs and provide structured correction feedback
Design A/B experiments: run auto-adjudication in shadow mode alongside manual process to measure accuracy
Review weekly variance reports and identify where the model is making systematic errors
Provide structured feedback records (Recommendation ID, Feasibility, Impact, Corrective Guidance) to retrain the model

Meta Interface Responsibility

Presents auto-adjudication accuracy metrics to Meta; negotiates rule refinements based on model performance data

Data Capture & Integration Specialist

Technology / Innovation

Specific Deliverables & Actions

Build and maintain the Plus–Jira bidirectional API integration (the #1 foundation deliverable)
Configure MCP connector to expose Plus discrepancy data to the AI model in standardized format
Set up camera systems at B2B receiving stations and dock bays — hardware installation + software config
Build the data pipeline that feeds scanned images, serial numbers, and manifest data to the vision model
Create the trailing 30/60/90-day response analytics data pipeline from Jira + Plus
Maintain the automated pull ticket generation system that routes to material handlers
Monitor data quality: ensure all scan events, photos, and API calls are logging correctly

Meta Interface Responsibility

Provides Meta with API documentation for the bidirectional sync; troubleshoots data flow issues

Operations & Process Analyst

Operations / Quality

Specific Deliverables & Actions

Map the current D2C receiving flow end-to-end: FedEx truck → dock → flex conveyor → scan → Plus → Problem Log
Identify every manual touchpoint that can be automated and quantify the time/cost of each
Define the secure unit routing protocol to eliminate un-wiped units in outside storage
Design the new CSR workflow: what manual steps remain after automation? Document the exception handling process
Train receiving operators and CSRs on the new automated workflow (reduced training burden is a key win)
Track cycle time metrics: time from discrepancy detection to resolution, before and after automation
Manage the aging case backlog cleanup in Plus (immediate action item)

Meta Interface Responsibility

Coordinates with Meta on disposition workflow changes; ensures Meta's response SLAs are met

Engineering & Quality Lead

Engineering / Quality

Specific Deliverables & Actions

Curate the historical scrap records dataset for vision model training — ensure data is labeled correctly
Define hazardous goods criteria for the vision model's scrap classification logic
Validate vision model accuracy: compare AI scrap/no-scrap decisions against manual expert assessments
Work with Meta to define Legal Hold vs. No Legal Hold determination criteria (open item #1)
Establish quality gates for auto-adjudication: what accuracy threshold must be met before going live?
Review B2B vision model outputs: are box labels being read correctly? Are quantity counts accurate?

Meta Interface Responsibility

Primary technical contact for Meta on scrap criteria, Legal Hold rules, and quality standards

Project Manager / Scrum Lead

Program Management

Specific Deliverables & Actions

Own the 4-phase implementation roadmap: sprint planning, dependency tracking, milestone reporting
Manage the 8 open items list — assign owners, track progress, escalate blockers
Coordinate Meta engagement: schedule review sessions, collect feedback, manage expectations
Run weekly standups with the Dream Team; produce status reports for executive sponsors
Track KPIs: aging reduction, cycle time improvement, auto-adjudication accuracy, labor savings
Manage the transition from shadow mode (Phase 3) to live auto-adjudication — change management
Ensure the local site team in Grapevine is trained and ready to execute the new workflows

Meta Interface Responsibility

Primary relationship manager with Meta program team; owns the communication cadence and escalation path

Open Items & Dependencies

#Open ItemOwnerPriorityBlocker
1Define Legal Hold vs. No Legal Hold determination criteriaMeta / RXT JointHighBlocks scrap auto-adjudication
2Confirm B2B box labels are consistently scannable across all suppliersRXT OperationsHighBlocks B2B vision pilot
3Obtain historical scrap records for vision model trainingMetaMedium
4Camera hardware selection and dock installation logisticsRXT FacilitiesMedium
5Plus API integration specifications for real-time Jira syncRXT IT / MetaHighPhase 1 dependency
6Obtain Meta's adjudication decision rules to enable auto-adjudicationMetaHighBlocks auto-adjudication
7Define secure storage protocol for un-wiped discrepancy unitsRXT Operations / SecurityHighCompliance risk
8Address aging discrepancy case backlog in PlusRXT CSR ManagementHighImmediate action

Reference Videos

AI Agents for Visual Inspection in Manufacturing

NVIDIA demonstrates how visual AI agents are transforming semiconductor manufacturing with automated defect detection and root cause analysis using vision foundation models.

Warehouse Automation with Computer Vision

Real-world examples of camera-based receiving, counting, and label scanning in logistics operations.