SkyClaim Watch Tower

Inspiration

€4 billion in flight compensation goes unclaimed every year under EC261/UK261. The legal framework has existed since 2004 — it's buried under industry friction. Incumbents take 35% commission, work one jurisdiction at a time, and rarely fight back when airlines deny claims with vague "extraordinary circumstances" language. Most passengers never realise they're entitled to up to €600 per disrupted flight.

I built SkyClaim Watchtower because modern agentic AI can translate, cite, and forensically annotate at zero marginal cost. Combined with MeDo's no-code agentic surface, an aviation rights advocate that matches a £300/hour solicitor becomes possible. Free. Multilingual. Cited. Forensic.

What it does

SkyClaim Watchtower is a live global aviation rights observatory powered by an agentic multi-plugin architecture.

Live experience. A real-time disruption radar shows aircraft from a four-source ADS-B cascade (adsb.lol → adsb.fi → airplanes.live → OpenSky), with current aviation news from MeDo's News plugin streaming alongside dynamic disruption markers extracted by an autonomous IATA-detection agent.

Multi-modal claim input — drop a boarding pass (MeDo OCR extracts everything), speak your claim (MeDo STT + TTS confirmation), or pick a demo case.

Agentic verdict. ERNIE 5.0 orchestrates a multi-step reasoning workflow as a single agentic pass: identifies the applicable regulation (EC261/UK261/US DOT), computes great-circle distance via Haversine, queries an embedded case-law knowledge base, and cross-references live NOAA METAR weather to test the airline's extraordinary-circumstances defence. Returns: regulation cited, compensation in three currencies via MeDo ExchangeRate, confidence score, structured caveats. A Web Audio tone confirms eligibility.

Multilingual demand letter. One toggle translates into the airline's home language — 13 languages via MeDo Translation, three seconds. French for Air France, Japanese for ANA, German for Lufthansa.

Letter actions: save as PDF, open in email (the airline's customer-relations contact is auto-resolved by an MeDo AI Search agent against an embedded directory of 35 carriers, with intelligent fallback), generate a shareable receipt card via MeDo Image Generation, or aggregate into a class-action group letter via Supabase realtime when multiple passengers claim the same flight.

Bluff Detector (/decode). Paste the airline's denial email. An autonomous forensic agent classifies every sentence (legitimate / misleading / legally-wrong / vague-burden-shifting / ambiguous) and annotates each against the case law that refutes it. "Operational reasons" → underlined on burden of proof. Technical-fault excuses → flagged. "Voucher full and final settlement" → struck through (statutory rights cannot be waived). The agent then drafts a counter-letter automatically.

AI Lawyer Roleplay. An adversarial agent simulates the airline's defence counsel across five rounds, invoking extraordinary-circumstances defences, technical excuses, and procedural objections. Scored on rebuttal strength against the embedded case-law knowledge base — preparation for an actual ADR submission or small-claims hearing.

Persistent case lookup — anonymous 8-character IDs, 4-milestone follow-up timeline (Day 0/14/30/60) guiding users from initial claim through ADR escalation to small-claims court.

How I built it

Zero custom backend code. Zero DevOps. Zero API keys for the core flow. Built entirely on MeDo's no-code agentic surface via 18 PRD-driven iterations.

14 integrations composed end-to-end

9 MeDo native plugins:

Plugin	Where it fires
Google Text Translation	Letter translation toggle on /letter
Web Search	Live verification card on /verdict
AI Search	Airline customer-relations contact resolution
The News	Live disruption headlines + dynamic-marker extraction
OCR	Boarding-pass image to structured fields
ExchangeRate	Multi-currency verdict display with ECB static fallback
Speech-to-Text	Voice claim input with field extraction
Text-to-Speech	Voice confirmation readback
Image Generation	Shareable receipt cards via Kling AI Lite

4 external live data sources (no API keys — community endpoints proxied via Edge Functions):

• adsb.lol → adsb.fi → airplanes.live → OpenSky — four-source ADS-B cascade with automatic failover
• NOAA Aviation Weather METAR — extraordinary-circumstances evidence

Infrastructure: Supabase Postgres + realtime channels + Edge Functions (opensky-proxy, metar-proxy, exchange-rate proxy, speech-to-text proxy) · Leaflet + CartoDB DarkMatter map · Web Audio API verdict tones · ERNIE 5.0 agentic reasoning core.

Embedded legal substrate (verified against EUR-Lex and legislation.gov.uk): EC261/UK261/US DOT eligibility logic with jurisdiction routing · €250/€400/€600 tiers by great-circle distance · 3-hour delay rule codified into UK261 by SI 2023/1404 · extraordinary-circumstances classification rules · EC261 reform caveat (June 2025 EU Council general approach) · US DOT correction (explicitly excludes the withdrawn September 2025 compensation rule).

Production-grade resilience. Every external source has automatic failover and graceful degradation — the system maintains consistent UX even under network or upstream-API stress, with telemetry on every fetch.

Challenges I ran into

MeDo Custom Plugin OpenAPI import was temporarily disabled during the hackathon — I restructured around MeDo's native plugin surface. Turned out to be a feature: nine native plugins composed end-to-end made the build more MeDo-native than custom wrappers would have been.

OpenSky geo-blocking: Supabase Edge couldn't reach opensky-network.org (HTTP 000 connection timeout). I architected a four-source ADS-B Exchange cascade with automatic failover plus a deterministic position-interpolation simulation that advances aircraft along true_track bearings every 5 seconds.

Citation hallucination: ERNIE 5.0 invents case names if asked to recall them. I embedded the entire case-law knowledge base directly in the system prompt rather than relying on the LLM's recall, then verified every citation against primary sources.

Chinese-character leakage: ERNIE occasionally produced Chinese placeholder text. I engineered the system prompt to forbid it and added post-generation audit steps.

Voice STT and ExchangeRate edge cases: Both diagnosed via MeDo's automated Quality Analysis agent. Voice required MediaRecorder format alignment with the STT plugin's expected MIME type; ExchangeRate needed structured error logging plus ECB static-rate fallback so the multi-currency row always renders.

Accomplishments I'm proud of

14 integrations composed end-to-end, zero custom code. Production-grade implementation depth matching enterprise SaaS submissions — built in a no-code chat surface.

Multi-agent orchestration as a single reasoning pass. ERNIE 5.0 coordinates jurisdiction detection, distance computation, case-law retrieval, weather cross-referencing, and confidence scoring as a unified agentic workflow — not a brittle chained prompt sequence.

Forensic AI for legal-text analysis (the Bluff Detector). An autonomous agent performing sentence-by-sentence case-law annotation of airline denials is novel for the consumer-rights vertical. No existing flight-compensation product does this.

Adversarial agentic roleplay (the AI Lawyer). A separate agent acting as opposing counsel — invoking the actual rhetorical patterns airlines use — is a credible legal-preparation tool, not a chatbot novelty.

Verified legal substrate. Every cited regulation and case verifiable against primary sources. The pending EC261 reform is acknowledged; the withdrawn US DOT rule explicitly excluded — precision a real aviation solicitor would recognise.

Multi-modal input: text, OCR boarding-pass extraction, and voice with STT + TTS confirmation — three input modalities composed in a single agentic surface.

Class-action coordination via Supabase realtime — collective leverage that no individual claim-letter generator offers.

What I learned

MeDo's native plugin composition is more powerful than the marketing implies. Nine plugins covered every external concern — I planned custom wrappers for AviationStack, PDFShift, Resend, and Companies House and needed none of them.

The PRD-first workflow is the right discipline for complex agentic apps. Each PRD update costs ~15 credits versus 80-130 for a full Deep Build regeneration — caught errors before burning the budget and produced a more coherent multi-page architecture.

Multi-agent orchestration scales better than prompt chaining. ERNIE 5.0 handles parallel agent tasks (jurisdiction + distance + case-law + weather + confidence) as a single reasoning pass rather than as sequential prompt invocations. Lower latency, lower cost, less error propagation.

The legal framework already exists. EC261 was written in 2004. UK261 codified in December 2023. Case law mature and unambiguous. The bottleneck has never been the law — it's been access. Agentic AI removes that bottleneck at near-zero marginal cost.

What's next for SkyClaim Watchtower

Jurisdiction expansion: Brazil ANAC Resolution 400/2016, Canada APPR, Australia ACCC, EU rail rights (Regulation 1371/2007) for multi-modal disruption claims.

Auto-submission integration to airline web-form APIs for carriers that no longer accept email (Ryanair, Wizz Air, IndiGo). Currently mailto:; next step is one-click submission against published airline endpoints.

Crowdsourced strike intelligence: aggregate passenger-reported airline responses to expose inconsistent extraordinary-circumstances defences across cases.

Class-action NEB escalation when claimant count exceeds 100 on a single flight — auto-generate a coordinated submission to the relevant National Enforcement Body for collective adjudication.

Lost-baggage claims under the Montreal Convention (up to ~£1,300 per passenger). Same legal-citation pattern, different regulatory framework.

Receipts mode: photograph all boarding passes throughout the year — the OCR pipeline proactively flags claimable disruptions the user would otherwise forget.

AI Tribunal: extended adversarial agentic roleplay that prepares users for actual ADR / CAA / small-claims hearings, scored against a learned model of which arguments succeed in each jurisdiction.

Mobile-native apps via Capacitor wrappers around the existing responsive H5 deployment, with push notifications when MeDo's News plugin detects disruption affecting a saved upcoming flight.

Built With

• medo