Commander files
5
Period list, ruby summary, department, fuel hose, cashier period
Demo data only — this is a sanitized sample, not a real store or live data. StationMind AI is an AI back-office platform in active development and pilot-readiness stage.
StationMind AI Demo StoreSanitized data
AI back-office that shows the owner what needs attention.
StationMind AI replaces 2-4 hours of daily back-office paperwork with review-gated AI decisions. Invoices, fuel reconciliation, daily close, and pricing — automated with guardrails, not blind trust.
AI decisions
14
Review-gated suggestions with source evidence
Owner actions
4
Only true exceptions are waiting
Unsafe automation
0
No price push, no POS write-back, no auto-finalized accounting
79+ routes
Full-featured Next.js 16 app with app router, covering every operational workflow
200+ tests
Vitest unit tests and Playwright E2E tests covering security, tenant isolation, and data flow
270+ lib files
Structured domain services for AI, connector, accounting, fuel, pricing, and operations
5,300-line schema
Prisma schema with 30+ enums, 50+ models, and strict tenant-scoped relations
Multi-tenant
Organization > Store hierarchy with privacy layers and audit-logged data access
Zero-trust AI
Every AI decision is review-gated, confidence-scored, and source-evidence-backed
Operating flow
What the demo proves
Connector
WorkingCommander data is arriving read-only
The demo shows direct CGILink-style report pulls, file receipts, heartbeat status, and archive training without showing credentials.
Evidence: Sanitized connector heartbeat and five sample Commander XML receipts
Review Inbox
Needs reviewOwner sees decisions, not raw system objects
Documents and POS signals become simple approval cards: what it is, where it goes, why AI thinks that, and what happens next.
Evidence: Four sample review cards with confidence, routing, and money impact
Daily Close
Needs reviewDraft close is ready but not finalized
StationMind AI can draft the close from POS-side evidence, then keeps final accounting locked until the owner approves missing or uncertain data.
Evidence: Department totals, cashier periods, fuel summary, missing EFT match warning
Fuel
Needs reviewPOS fuel sales are visible; tank/SIR remains pending
Commander hose data is treated as POS sales truth only. Certified fuel reconciliation waits for fuel invoice and tank/SIR/ATG evidence.
Evidence: Sample hose report, invoice-pending status, tank/SIR missing exception
Ordering
Safety disabledSuggestions only, no auto-order
The demo surfaces fast movers, missing counts, and invoice-driven reorder clues while keeping purchase orders approval-gated.
Evidence: Three sample reorder suggestions, no vendor submission
Safety
WorkingProduction guardrails are visible
Read-only Commander integration, secret redaction, tenant isolation, audit events, and high-risk approval requirements are shown as first-class product behavior.
Evidence: Safety checklist with disabled write-back and disabled price push
Review Inbox
Business decision cards
Each card shows what the item is, where it goes, why AI thinks that, and what money/report it affects. The owner approves or rejects — StationMind AI never auto-finalizes.
Core-Mark invoice
Core-Mark
Amount
$2,184.42
Destination
Purchase / Inventory
Known vendor and invoice pattern; new tobacco item mapping needs owner confirmation
Updates vendor spend and inventory cost basis after approval
Confidence 91%
Fuel delivery invoice
Fuel supplier
Amount
$18,460.17
Destination
Fuel Purchase
Fuel invoice detected, but grade mapping needs review
Feeds fuel inventory and margin draft; not compliance-certified without tank/SIR
Confidence 86%
Card settlement notice
Processor email
Amount
$6,932.28
Destination
Credit Card Settlement
External deposit source is present, but fee split is not certified
Matches POS expected card totals to deposit candidate
Confidence 79%
AT&T bill
AT&T
Amount
$244.18
Destination
Telecom Expense
Recurring sender and vendor, low payment-risk language
Updates expense report and P&L draft after approval
Confidence 94%
Impact
Before vs. After StationMind AI
Sample data — these represent the target workflow improvements, not measured production metrics.
Invoice Processing
Before
Owner manually opens emails, identifies vendor, enters line items into spreadsheet — 20-30 min per invoice
After
StationMind AI classifies the document, extracts fields with OCR, routes to correct account, and presents a one-click approval card
Daily Close
Before
Owner manually cross-references POS register report, cashier periods, and card settlements — error-prone, 45-60 min daily
After
StationMind AI drafts the close from POS evidence, flags discrepancies, and waits for owner approval before finalizing anything
Fuel Operations
Before
Owner manually tracks fuel deliveries, hose readings, and tank dips on paper or Excel — compliance gaps are invisible
After
StationMind AI ingests hose data automatically, flags missing tank/SIR evidence, and refuses to generate compliance reports without proper proof
Pricing Updates
Before
Owner tracks costs in head, updates shelf labels based on gut feel, misses margin erosion until month-end P&L
After
PricePilot surfaces cost changes from invoices, recommends margin-aware prices, and keeps all changes review-gated via StationMind AI
AI Explainability
What the AI does
Every AI capability is review-gated and confidence-scored. The owner sees inputs, outputs, and reasoning. StationMind AI uses AI as a tool under human supervision, not as an autonomous agent.
Document Classification
RealRoutes incoming documents (invoices, fuel deliveries, EFT statements, expenses) to the correct accounting category
Inputs: OCR text, sender metadata, document structure
Outputs: Category, confidence score, routing destination
Vendor Recognition
In ProgressIdentifies vendors from invoice patterns, maps line items to products, learns from corrections
Inputs: Invoice text, historical vendor patterns, product catalog
Outputs: Vendor match, item mappings, confidence per field
Pricing Recommendations
In ProgressSuggests price changes based on cost changes, competitor data, and margin targets
Inputs: Cost basis from invoices, margin rules, competitor prices
Outputs: Recommended prices with explanation and margin impact
Fuel Reconciliation
In ProgressCross-references POS hose data with delivery invoices, tank readings, and ATG data
Inputs: Commander hose reports, fuel invoices, tank/SIR readings
Outputs: Variance reports, missing evidence flags, exception cards
Daily Close Automation
In ProgressDrafts end-of-day accounting from POS evidence, flags gaps, keeps finalization owner-gated
Inputs: Cashier periods, department totals, card settlements, fuel summary
Outputs: Draft close sheet, missing data warnings, approval queue
Reorder Intelligence
In ProgressIdentifies fast-moving products, low stock signals, and vendor deal opportunities
Inputs: Sales velocity, invoice history, vendor catalog, stock estimates
Outputs: Reorder suggestions with urgency and savings potential
System design
Architecture
Frontend
Next.js 16 app router with 79+ routes, React 19, Tailwind CSS 4, dark/light mode
TypeScript, React 19, Next.js 16, Tailwind CSS 4, TanStack Table
API & Auth
NextAuth 4 with role-based access, tenant-scoped middleware, and audit logging
NextAuth 4, Zod validation, middleware route enrichment
Database
PostgreSQL with Prisma ORM, 30+ enums, 50+ models, strict tenant isolation
PostgreSQL, Prisma 6, tenant-scoped queries
AI Pipeline
Document OCR, classification, confidence scoring, learning feedback loops
LLM provider abstraction, Zod-validated extraction, trust engine
Connector
Read-only Verifone/Commander integration with heartbeat, evidence archiving
Windows service, XML/CSV parsing, S3 archiving
Job System
BullMQ background processing for document ingestion, report parsing, training
BullMQ, Redis, worker runtime, scheduled jobs
Development
Roadmap
Phase 1: Foundation
Completed- Multi-tenant auth and store onboarding
- Commander connector (read-only)
- Document ingestion pipeline
- Review inbox with decision cards
- Daily close draft from POS evidence
Phase 2: Intelligence
In Progress- AI document classification and routing
- Vendor recognition with learning loops
- Fuel operations and hose data parsing
- Pricing rule engine and recommendations
- Confidence scoring and trust levels
Phase 3: Scale
Planned- Multi-store regional management
- Automated reorder suggestions
- Bank reconciliation automation
- Mobile-first operator workflows
- Certified fuel compliance reporting
Responsible AI
Safety-first design
No silent automation
Every AI decision requires explicit owner approval. StationMind AI will not push prices to the POS, auto-finalize accounting, or place vendor orders without the owner clicking Approve.
Confidence transparency
Every classification, routing, and extraction decision shows a confidence score. Low-confidence items are flagged for review. The owner can see why AI made each decision.
Learning from corrections
When the owner corrects AI, the system captures the correction, stores it in the learning pool, and improves future classifications — but only within the privacy-safe learning boundary.
Tenant-isolated data
Store A's data never influences Store B's AI. Learning happens in privacy layers: store-private corrections stay private, only anonymized patterns reach the platform pool.
Build story
How this was built
StationMind AI is built by a solo founder using AI-assisted development tools (Claude, Copilot) alongside traditional engineering. AI tools accelerate code generation, pattern implementation, and documentation — but every architectural decision, security boundary, and business rule was designed and verified by a human engineer.
The platform is currently in development and demo/pilot-readiness stage. The demo uses sanitized sample data to show the platform's capabilities without exposing real store operations. Features labeled “Real” have production-ready code. “Demo” features work with sample data. “In Progress” features are actively being built.
Resume bullets
- Built a full-stack AI back-office platform for gas station and convenience store owners, handling document classification, pricing, fuel operations, and daily accounting.
- Designed a review-gated AI system where the owner approves every business decision — no silent automation, no unsupervised price changes, no auto-finalized accounting.
- Implemented a read-only Verifone/Commander connector with source evidence archiving, heartbeat monitoring, and multi-format report parsing (XML, CSV, PDF).
- Built a multi-tenant SaaS architecture with strict tenant isolation, privacy layers (store-private, org-private, platform-safe), and audit logging on every data access.
- Created an AI confidence scoring system with trust levels, learning feedback loops, and human correction capture for continuous model improvement.
- Engineered document OCR pipeline with vendor-specific extraction templates, confidence-scored field mapping, and exception routing for uncertain classifications.
Interview walkthrough
- 1Start with the problem: gas station owners spend 2-4 hours daily on back-office paperwork — invoices, fuel reconciliation, daily close, pricing — that could be automated with the right AI guardrails.
- 2Show the Review Inbox: every AI decision becomes a simple card the owner can understand — what it is, where it goes, why AI thinks that, and what money it affects. The owner stays in control.
- 3Walk through the Connector: PricePilot reads POS data from the store's Commander system in read-only mode. No write-back, no price push. The connector proves it's working with heartbeat + file receipts.
- 4Explain the trust model: AI confidence scoring determines what routes automatically vs. what needs owner review. New vendors always go to review. Known patterns with high confidence can be pre-approved.
- 5Show fuel operations: POS hose data is treated as sales truth only. Certified fuel reconciliation waits for external evidence — fuel invoice, tank/SIR, ATG readings. PricePilot won't fake compliance.
- 6Close with safety: tenant isolation, secret redaction, audit logging, and the principle that no accounting is ever finalized without explicit owner approval.