Mustafa.
ende
Back to all projects

Urla Shoes — AI document-compliance engine + multi-feature Salesforce platform

A multi-feature Salesforce build led by an AI Partner Compliance Document Manager: an Einstein Prompt Template (Prompt Builder, GPT-4o mini) classifies each uploaded partner document into structured JSON, validates it, and auto-closes the matching compliance request. Around it: route-safety AI (Google Maps + OpenWeather + Einstein), Contact enrichment, Lead Queue routing, Loan Sync, a Reseller Matching engine, and a Custom-Setting-backed API key layer that keeps secrets out of source.

Year
2024
Role
Salesforce Developer
Tech stack
Apex (Queueable, Schedulable)Lightning Web ComponentsVisualforce (bridge)Salesforce EinsteinEinstein Prompt Templates (Prompt Builder)AI document classification (GenAI)Agentforce actions (@InvocableMethod)Notion APICustom Metadata TypesConnectApi.EinsteinLLMTrigger Framework (Kevin O'Hara)Custom Settings (hierarchical)Platform EventsCustom ObjectsHttpCalloutMockRemote Site Settings

Highlights

  • AI Partner Compliance Document Manager — an Einstein Prompt Template (Prompt Builder, GPT-4o mini via ConnectApi.EinsteinLLM, Trust Layer) classifies each upload into structured JSON (category, confidence, expiry), validates it, writes it back, and auto-closes the matching document request
  • Compliance Cockpit — a live MISSING/EXPIRED/EXPIRING/OK matrix with per-partner scores, driven by Custom Metadata requirements (Tax Certificate, Insurance, Signed Contract, Bank Details)
  • Agentforce invocable actions (@InvocableMethod) — LeadQualificationService scores + rates leads (Hot/Warm/Cold) with a DACH bonus (DE/AT/CH) and auto-creates a call task; OnboardReseller derives tier (Platinum/Gold/Silver) + commission and spins up Account + Contact + Opportunity + Quote
  • Magic-link partner self-registration — unmatched resellers get a single-use, 7-day tokenized link; an @InvocableMethod reconciliation action later back-links them to every unmatched opportunity by email domain
  • Notion API integration — programmatic STAR-format pages (block-JSON, language-token alias map, 2000-char rich-text chunking) from a protected Custom Setting token
  • Route Safety LWC — Google Maps + 5-waypoint parallel OpenWeather + Einstein Prompt Template AI verdict in one screen
  • Contact Enrichment Queueable — Nationalize.io with HttpCalloutMock across 6 scenarios (success, empty, HTTP 500, missing FirstName, bulk-10, parser test)
  • Shift-based lead-routing engine — round-robin by lowest load with per-shift capacity, 5-min SLA deadlines that reclaim breaches (SLA_Breach_Log__c) and re-distribute excluding the breaching reps, timezone/DST-correct windows incl. overnight shifts; scheduled batch + Lead_Shift_Event__e Platform Event + LWC dashboard
  • Loan Sync Pipeline — Loan__c + Loan_Sync_Log__c custom objects + Opportunity-triggered sync + retry scheduler with audit logging
  • Reseller Matching Engine — multi-criteria scoring + Opportunity trigger + LWC tier badge
  • Platform foundations — extensible Kevin O'Hara TriggerHandler framework, round-robin task assignment, SLA expiry automation
  • Senior pattern: API keys served from API_Config__c Custom Setting via ApiKeyService Apex helper — no secrets in source

Feature walkthrough

Partner Compliance Cockpit

A network-wide matrix scores every reseller's required documents at a glance — MISSING / EXPIRED / EXPIRING / OK — with per-partner compliance percentages and roll-up counts (fully compliant, expiring soon, at risk). The required documents are defined in Custom Metadata, so an admin can change the ruleset without a code deploy.

Partner Compliance Cockpit — reseller network compliance matrix with per-partner scores

AI Document Classifier (Einstein / GPT-4o mini)

Each uploaded document is analysed on demand: an Einstein Prompt Template classifies it into a category with a confidence score, extracts its expiry date, and writes a one-line summary — then applies the result to the record. Here the model files a distribution contract (95%), a tax certificate (90%) and an insurance certificate (90%) automatically.

AI Document Classifier — Einstein classifying uploaded partner documents into categories with confidence scores

Live scoring & auto-fulfilled document requests

On each reseller record, a scored checklist reflects document status in real time via Lightning Message Service — no page reload. 'Request Missing Documents' raises a Document_Request__c and emails the partner's primary contact; once the matching document is classified, the request auto-closes and the score climbs from 25% to 100%.

Reseller compliance checklist reaching 100% with auto-fulfilled document requests

File Hub — Partner Document Center

A business layer over Salesforce's native ContentDocument storage that adds sender, recipient, category and record-link semantics, with colour-coded category tiles and a monthly activity heatmap. Private sharing via Apex-managed shares, ghost-record protection on deletion, and FLS/CRUD enforced with WITH USER_MODE (~91% controller coverage).

File Hub — partner document center with category tiles and activity heatmap

Shift-based lead-routing engine

Inbound leads are routed to reps in active, timezone-aware shifts using capacity-aware round-robin (lowest load wins). A 5-minute SLA reclaims un-actioned leads, logs an SLA_Breach_Log__c audit record and re-distributes them while excluding the breaching rep. A Platform Event feeds the live dashboard — a 24-hour global timeline and a champion board — with no polling.

Lead Shift Dashboard — 24-hour global coverage timeline and champion board

Loan creation & sync pipeline

A Closed-Won opportunity mints a Loan record and syncs it to a core system via a Queueable HTTP callout — one callout per batch even on bulk inserts. Every attempt is logged to Loan_Sync_Log__c with the request/response payload; a retry ceiling plus a 'Retry Now' action make it resilient and auditable.

Loan Sync Dashboard — sync status, JSON payload and retry history

Route & Weather Navigator (AI safety verdict)

A Google Maps driving route (Essen → destination) with live weather at five interpolated waypoints, ending in a three-sentence Einstein AI safety verdict (safe / caution / do not depart). Google Maps runs inside a Visualforce page bridged to the LWC via postMessage — the standard workaround for Lightning Web Security.

Route Safety LWC — Essen to Berlin route, 5-waypoint weather, Einstein AI safety verdict

The problem

I needed a single Salesforce project that demonstrated platform breadth — multiple integration patterns, async Apex flavours, AI integration, custom data models, and the kind of secrets-out-of-source discipline that production orgs require. A single-feature demo wouldn't show how engineering decisions compose across an org.

Architecture

The flagship is an AI Partner Compliance Document Manager. When a partner uploads a file to the File Hub, DocumentIntelligenceService builds a context string (file name, type, sender note, reseller, the live category taxonomy and which categories carry an expiry) and invokes an Einstein Prompt Template ('DocumentClassification', GPT-4o mini) through the Trust Layer via ConnectApi.EinsteinLLM.generateMessagesForPromptTemplate. The model returns strict JSON — category, confidence (0–100), an extracted expiry date, and a summary — which the service validates against the allowed picklist, writes back onto the File_Hub_Entry__c, and uses to auto-close any matching Document_Request__c; compliance requirements live in Custom Metadata (Tax Certificate, Insurance, Signed Contract, Bank Details), so ComplianceService computes a live MISSING/EXPIRED/EXPIRING/OK matrix and a per-partner score (the Compliance Cockpit), all FLS/CRUD-enforced (WITH USER_MODE) with @TestVisible LLM mocking. A second AI feature is the routeWeather LWC — a three-integration composition. It hosts a Visualforce page (RouteMapPage) in an iframe (the only way to load Google Maps under Lightning Web Security) and communicates via postMessage: the LWC sends DRAW_ROUTE, the page responds with ROUTE_DONE plus the destination coordinates. The LWC then computes five waypoints (origin, 25 %, midpoint, 75 %, destination) and fires five parallel fetches to OpenWeatherMap via Promise.all. Once weather lands, the LWC POSTs a slim six-field-per-waypoint payload to the Apex RouteWeatherAnalysisService, which formats the data as text and invokes ConnectApi.EinsteinLLM.generateMessagesForPromptTemplate('RouteWeatherAnalysis', input). The GPT-4o mini response is rendered in a colour-coded card by a keyword classifier (storm/ice/hazard → red; caution/careful → amber; otherwise green). Around that headline sits a Lead Queue routing system (LeadQueueScheduler scheduled batch + Lead_Shift_Event__e Platform Event + leadShiftDashboard LWC), a Loan Sync Pipeline (Loan__c + Loan_Sync_Log__c custom objects with retry scheduler and audit logging), a Reseller Matching Engine (multi-criteria scoring service + Opportunity trigger + resellerTierBadge LWC), and an extensible TriggerHandler framework that every trigger in the org extends.

The approach

Three principles run through every feature. First, async by default — every external callout uses Queueable or Schedulable, no synchronous trigger callouts, no governor surprises under load. Second, secrets out of source — both the Google Maps and OpenWeather API keys live in an API_Config__c hierarchical Custom Setting, read at runtime via an ApiKeyService Apex class (instance property for VF, @AuraEnabled cacheable methods for LWC). The Visualforce page renders the Maps key via {!JSENCODE(mapsApiKey)} so it never appears as a literal in source; the LWC awaits getOpenWeatherApiKey() in connectedCallback. ApiKeyServiceTest gives 100 % coverage. Third, every trigger extends the same Kevin O'Hara TriggerHandler base — recursion guards, bypass mechanism, testable logic — so the routing, sync, matching, and contact-enrichment features all behave consistently under bulk load.

The outcome

A reference Salesforce sandbox built end-to-end (private repo; walkthrough on request). Its flagship is an autonomous AI document-compliance loop (Einstein Prompt Templates / Prompt Builder, structured-JSON classification, auto-fulfilled requests); around it sit route-safety AI, Contact enrichment, Lead Queue routing, Loan Sync, and a Reseller Matching engine — across LWC, Apex, Visualforce, custom objects, Custom Metadata, Platform Events, scheduled batches, and Einstein AI. The senior-engineering details (key management via Custom Setting, trigger framework, async-by-default, FLS/CRUD with USER_MODE, mocked LLM tests) are visible in the diff. The same patterns transfer directly to the multi-feature Salesforce orgs that mid-to-large DACH companies actually run.

External integrations

GitHub & Notion — private, walkthrough on request