Dillon Green
ALL WORK

CASE STUDY Internal tool · sole developer

PAR: parts, orders and repairs for a flight simulator fleet

PAR is the system rolled out to an airline's 40-person flight simulator maintenance department to track parts, order them, repair the simulators they belong to and account for the money. It went live in October 2026, and most of what's been added since came from the people using it. I designed it, built it with Claude Code, and run it.

Codebase
~100klines incl. tests, one developer
Surface
322routes · 3 roles
Schema
71SQLite tables + audit ledger
Tests
1,900+automated checks
In daily use by
40person department
Fleet
26simulator devices, two sites
Migration
20+importers, legacy and ongoing
Schematics
6,000+drawings converted to SVG

01

Problem

A full flight simulator is an aircraft cockpit plus racks of computers, I/O and motion hardware, and it has to pass regulatory qualification to stay in service. Keeping two dozen of them in that state means always knowing which part is where, what condition it's in, what's on order and which budget is paying for it.

The department's records were split between a legacy MS Access database and GigaTrak, and the two had drifted apart. Part conditions, the work to fix them, purchasing and budgets weren't connected in one place. The schematics that show where a part sits in a simulator were MicroStation CAD drawings, which needed CAD software to open and weren't linked to stock.

The department needed one system of record that technicians would actually use on the floor, and a migration that didn't lose history.

02

What I built

  • Inventory that matches the hangar

    Consumable, rotable and repairable parts, serial tracking, and a building → location → bin hierarchy.

  • Requisition to shelf

    Requisitions become purchase orders, and receiving an order puts the parts straight into inventory.

  • Work orders that open themselves

    When a part's condition changes, PAR opens the work order to repair it, so nothing broken gets lost on a shelf.

  • Budgets and project accounting

    Spending is tracked against budgets and projects in the same system as the parts and orders.

  • Fleet-level reorder logic

    Spares are aggregated across 26 simulator devices at two sites, with critical and low-stock tiers and ABC/FSN classification.

  • Schematics linked to stock

    6,000+ CAE MicroStation DGN drawings converted to SVG through GDAL, with part numbers on the drawing linked to live stock.

  • A migration that keeps history

    20+ importers, most of them off MS Access and GigaTrak, including a reviewable two-way, field-level reconcile engine that learns vocabulary mappings by majority vote, and an idempotent history import that keeps original timestamps.

  • Integrations on the floor

    QR labels in Zebra ZPL through a network print agent, photo capture and barcode scanning from phones, and a keyed REST API written for the discrepancy system.

03

Since go-live

PAR has a feedback button on every page, and most of this list started as a note from someone on the floor or in the office.

  • Store-room kiosk

    An iPad in the store room runs a part finder. Sign in with a PIN, find the part, take it, and PAR records who has it.

  • Scanning and shelf counts

    A scan page for barcode scanners and phone cameras, shelf counts, and moving a whole basket of parts in one all-or-nothing step.

  • Calibration tracking

    Due dates and certificates for calibrated tools and facility equipment, brought in from a separate tracker.

  • Shipments between buildings

    Internal shipments get an in-transit status and notify the sender and the requester, so a part on a truck isn't a missing part.

  • Approved alternates

    Form-fit-function alternates count as one part number for low stock and reorder, with a check page for groups that don't line up.

  • Rotable history

    Time installed, on the shelf and out for repair for each rotable part, worked out from the transaction history, with shelf-life flags.

  • Paperwork on the part

    Release certificates and other documents attach to the part they belong to, and parts link to the requisition that bought them.

  • Card purchases on the budget

    Purchasing-card statement exports import as closed requisitions, so card spend lands on the same budget lines as purchase orders.

  • Reports and procedures

    A weekly parts report as a PDF, and 11 step-by-step procedures with screenshots taken from a scrubbed database.

  • The old records, still readable

    The legacy system's fault reports are kept as a browsable archive with a timeline per serial number.

04

Architecture

PAR is deliberately boring to operate. It's one Python process: Flask behind Waitress, server-rendered Jinja templates with Bootstrap, and SQLite in WAL mode, with no build step on the front end. It runs as a Windows service from a PyInstaller bundle. A deploy is a folder of the files that changed plus a build stamp, copied over and restarted, and the About page shows which build is running.

For a department of this size, a single-file database with WAL gives concurrent reads and very simple backups. A 30-thread concurrency test keeps that assumption honest. Three roles control what each person can change, and an immutable audit ledger records every change.

PAR architecture Browsers, phones, a store-room kiosk and the discrepancy system's REST API call a Flask application served by Waitress, which stores data in SQLite. The app runs as one Windows service process. Importers bring in data from MS Access, GigaTrak and CAE documentation. A schematics pipeline converts DGN drawings to SVG with GDAL. The app sends labels to a network print agent, and the database is backed up daily and offsite. CLIENTS · FLOOR DATA IN · OUT solid: requests, data in dashed: output leaving WINDOWS SERVICE · ONE PROCESS Desktop browsers Jinja + Bootstrap pages Phones · kiosk iPad scan · photos · part finder Discrepancy system keyed REST API Waitress · Flask 322 routes · 3 roles server-rendered Jinja REST API · print jobs SQLite (WAL) 71 tables immutable audit ledger scrubbed copies for agents Print agent ZPL QR labels → Zebra 20+ importers Access · GigaTrak · CAE docs Schematics pipeline DGN → GDAL → SVG Backups daily + offsite PAR architecture: clients call a Flask app on Waitress backed by SQLite, fed by importers and a schematics pipeline, with a print agent and backups. CLIENTS Browsers Jinja pages Phones scan · kiosk Discrepancy REST API Waitress · Flask 322 routes · 3 roles server-rendered Jinja REST API · print jobs SQLite (WAL) 71 tables immutable audit ledger scrubbed copies for agents WINDOWS SERVICE · ONE PROCESS DATA IN 20+ importers Access·GigaTrak Schematics DGN→GDAL→SVG DATA OUT Print agent ZPL → Zebra Backups daily + offsite solid: requests and data coming in dashed: output leaving the service
Simplified. Names of internal systems, sites and suppliers are left out on purpose.

05

Engineering practices

PAR is built with Claude Code, so the tests matter: 1,900+ automated checks run before every deploy package and catch most regressions before people on the floor do.

140
Regression suites

Run in parallel, each against its own copy of the database. About 1,800 named checks.

171
Browser end-to-end checks

Headless Chrome, driven by a small, dependency-free Chrome DevTools Protocol client. A second crawl measures text contrast in all four themes.

30
Threads, one test

A concurrency test that hammers the app and SQLite at once.

The quality gate ratchets: the bar can go up, never down. Backups run daily and offsite. Agents work from counts, not records: when they need to look at data they get a scrubbed copy, and test suites run against throwaway copies that only report pass/fail and counts.

06

Results

Live since October 2026 and in daily use by the whole 40-person department. Inventory, purchasing, work orders, budgets and schematics now live in one system instead of several that disagreed.

The legacy MS Access and GigaTrak data came across through imports a person could review, field by field, without losing the original history or timestamps.

Technicians can open a simulator drawing in the browser, click a part number and see what's on the shelf. Reorder decisions are made across all 26 devices at both sites instead of one device at a time.

It's still changing. Feedback comes in through the app, and fixes go out as small packages with the tests run first.

Internal tool — details generalized; no proprietary data shown.