2026-09-30 · 5 min read

How a Non-Engineer Built FlightClaim

The first version of FlightClaim was a beautiful lie.

It had a landing page that looked like a Series B startup: a hero, a gradient, testimonials from "Deloitte" and "Oracle," a stat that said $2.4M recovered and another that said 94% success rate. None of it was real. I hadn't recovered a dollar for anyone. The "Check my flight" button went nowhere. I'd asked an AI to build me a flight-compensation product and it had, faithfully, built me the appearance of one: 800 lines of polished nothing.

I sat with that brochure for a while. Then I deleted it and started over, with a rule: the product is the part that's hard, not the part that's pretty.

The problem: a rulebook designed to be ignored

When a flight is delayed three hours or more, cancelled on short notice, or overbooked, the airline often owes you a fixed amount of cash. Not a voucher, not miles. Money, up to €600 in the EU and up to £520 in the UK, set by regulation and indexed to how far you were flying.

Almost nobody collects. Three things stack against you.

• The rules are opaque. Whether you're owed depends on distance bands, the length of the delay at your final destination, how much notice you got, and whether the cause counts as an "extraordinary circumstance." Airlines are not eager to explain this.
• Denial is the default. The airline's cheapest move is to say no and wait. Most people stop there.
• The middlemen skim. Claims-management firms will take the case off your hands and keep a quarter to half of the payout to send a demand letter a passenger could send for free.

That last one is the tell. An industry monetizes the gap between what the law says and what a passenger can figure out. Close the gap and the tax disappears.

What I built

FlightClaim asks five questions (route, airline, what went wrong, and the few follow-ups the law turns on) and gives you a verdict in under a minute.

• A yes, a no, or a likely. Not a maybe. With the exact amount you're owed.
• The reasoning, shown. Every verdict cites the rule it rests on, so you can see why, not just what.
• A ready-to-send demand letter, regulation cited, deadline set, pre-filled, to copy or download.
• Honest no's. If you're not covered it says so, and points you to the right regime instead: US Department of Transportation refund rules, the Montreal Convention. No false hope.

The verdict is computed in your browser. Your itinerary never leaves the page.

The layer below: teaching an app to read case law

The core of FlightClaim isn't the interface. It's a rules engine, a few hundred lines of pure logic that encode how European courts have actually ruled on flight compensation. Not my opinions about the law. The law.

When the app decides whether your delay was an "extraordinary circumstance," the airline's main escape hatch, it isn't guessing. It checks a table that maps each cause to a verdict, and each verdict to a real case.

• A technical fault is not extraordinary (Wallentin-Hermann, C-549/07).
• A wildcat strike by the airline's own staff is not extraordinary (Krüsemann, C-195/17). An external air-traffic-control strike is.
• Cause unknown? The table returns "likely eligible" and puts the burden on the airline, which is where the law puts it too.

There's more of this than I expected. Compensation for a connecting flight is measured from where you ended up, on one booking, by direct distance rather than the sum of the legs: three rulings braided together (Folkerts, Bossen, Wegener). The post-Brexit seam, where a UK carrier flying into the EU becomes a "third-country carrier" and the jurisdiction flips, is the exact spot most tools get wrong.

Once I saw that the maze could be a table, cause in one column, verdict and citation in the others, the project stopped being a legal problem and became a data problem. A year ago I couldn't have built it. What I could do was read the regulation, find the cases, decide how the logic should branch, and work with AI to turn that into an engine with seventy passing tests. The judgment (which cases apply, how to weigh an unknown cause, when to say "not covered" instead of inventing hope) was mine. The typing was cheap.

What building it taught me

The honest version beats the miracle version, and it's the only one you can ship. From my project notes: "The tempting version of this product promises a miracle: connect your inbox, we watch your flights, money appears. It reads well and does nothing until every hard part works. The honest version is smaller and more useful." The miracle version is a demo. The honest version is a tool.

AI will happily build you the brochure. V1 wasn't a failure of the model. It was a failure of my brief. I asked for a flight-compensation product and got the surface of one, complete with invented proof. The lesson isn't that AI lies. It's that AI gives you exactly the shape you specify, so the specifying is now the craft. Ask for the hard part first.

The general method wins. A claims firm's edge was a spreadsheet of rules in a lawyer's head and a template letter. Encode the rules once and the edge is gone. It's the same move I described in The Layer Below: the hand-held, specialized process loses to the general, computable one.

The tension I haven't resolved

Honesty cuts both ways, so here's mine.

I built FlightClaim to remove a middleman that keeps a third of your payout for sending a letter. Then I put that same letter, along with the escalation kit, behind a one-time unlock and an account. The verdict is still free. The letter is not. I can defend it: the reasoning that produces the verdict is the expensive part, and a flat fee of a few euros is not 35% of €600. But I won't pretend it sits perfectly clean against the "asks for nothing back" version I started with.

What would make the whole project wrong? If the engine is confidently incorrect. A cited, official-looking verdict that's wrong sends someone to argue a losing case with false confidence. That's why the rules engine is the part under test, and why every "extraordinary" call is pinned to a ruling I can point to. A demand letter built on a hallucinated regulation is exactly the failure I built V1 to stop making.

Press send

The rulebook was never the airline's secret. It was written to be tiring. FlightClaim takes the part they count on you not understanding and makes it legible in a minute, and then asks you to do the one thing no tool can do for you.

You press send. That's the whole product, and it's enough.