Project Overview

Steve is an image-based submission platform built for Open Loyalty workflows. Customers submit evidence in the mobile app, reviewers process it in the admin panel, and the Convex backend coordinates storage, AI analysis, fraud checks, review state, and optional Open Loyalty sync.

Primary actors

Actor	Surface	Responsibility
Customer	Mobile app	Capture evidence, submit workflows, track status, review results
Reviewer	Admin panel	Claim work, inspect files, resolve fraud, approve or reject
Admin	Admin panel	Reviewer capabilities plus company and customer management
Super admin	Admin panel	Workflow builder, AI config, fraud policies, API keys, workflow agent
External integrator	HTTP API	API-key-backed server-to-server integration

Product surfaces

Admin panel

Current admin capabilities include:

dashboard and inbox
company and customer management
fraud policy configuration
image moderation configuration
API key management
reporting and usage views
workflow list and workflow builder
unified submission queue
submission review detail
trash and restore flow

Mobile app

The mobile app is the customer-side capture client. Today it exposes four workflow entrypoints backed by the unified submission system:

document-ocr
compliance-vm
tire-replacement
roof-cleaning-inspection

Backend

The backend provides:

role-aware auth and profile separation
workflow CRUD and versioning
submission creation, review, and audit history
optional image moderation pre-check before workflow processing
AI analysis with configurable providers and fallback models
fraud detection using perceptual hashes and extracted-data comparison
Open Loyalty schema push and event sync
API keys, API sessions, webhook signing, and a partial external HTTP surface

How the platform is modeled

Steve is built around a generic workflow engine rather than separate hardcoded "document" and "compliance" backends:

workflow
  -> versioned configuration
  -> submission
     -> files
     -> events
     -> fraud matches

Each workflow version stores six stages:

Upload
Enhancement
AI analysis
Fraud detection
Review
Sync

That makes the same backend usable for OCR, store-compliance checks, and before/after verification flows.

A separate global image moderation pre-step can run before those stages. It is configured outside workflow versions because it is intended to protect all submission flows consistently.

Submission lifecycle

created -> processing -> review -> approved -> synced
                     \-> failed

Reviewers can also:

claim or unclaim review items
edit extracted data
resolve fraud findings
reprocess failed or review-state submissions
retry sync from approved state
cancel active work
move submissions to trash and restore them later

Current integrations

Integration	Purpose
Convex Auth	Auth and sessions
Cloudflare R2	Direct file storage and presigned uploads
OpenRouter	Primary AI provider for workflow analysis and image moderation
Resend	Invitation emails
Open Loyalty	Custom event schema push and event sync

Important current constraint

The repository contains a larger design for a full upload-session API, but the live external HTTP surface is still smaller than that design.

Live /api/v1 routes today:

GET /api/v1/companies
GET /api/v1/jobs/{sessionId}
POST /api/v1/webhooks/test

Planned upload-session routes are documented separately under Upload Session API.