Runway Sentinel — Airfield UAS Inspection Ops · KBOS

A drone fleet operations platform for airfield inspection at Boston Logan International (KBOS): runway FOD sweeps, edge-light checks, and perimeter patrol — from natural-language mission requests through Part 107 preflight compliance to findings and maintenance work orders.

Built for the HTCJ × PROOF Aviation Futures Challenge.

Live demo: https://runway-sentinel.vercel.app

What it does

• Mission planner — pick a mission type (FOD sweep, edge-light check, perimeter patrol), a target runway, and an altitude. The planner lays down a waypoint pattern over real KBOS geometry: every runway centerline/threshold comes from the public-domain OurAirports dataset, so pass spacing, distance, and timing are computed from true runway dimensions.
• Coverage math — altitude and camera FOV determine ground swath; passes are spaced for 20% sidelap so the stated ground-sample distance (cm/px) is honest, not illustrative.
• Fleet assignment — ranks airframes by mission fit: required sensor, wind limits vs. live conditions, battery state-of-health (multi-pack sorties with pack swaps when needed), and maintenance intervals.
• Preflight compliance engine — a real go/no-go checklist: Class B authorization reference (14 CFR 107.41 — KBOS sits inside the B surface area where LAANC grids read 0 ft), altitude vs. authorized ceiling, runway-closure/ATC coordination, wind vs. airframe limit, 3 SM visibility and 500-ft cloud clearance (107.51), daylight vs. civil twilight vs. night anti-collision lighting (107.29, computed by a solar algorithm verified against US Naval Observatory times), Remote ID (Part 89), maintenance interval, and 25% battery reserve.
• Live weather — the server proxies the real KBOS METAR from aviationweather.gov. Gusty days genuinely ground the quads (watch RANGER-4 become the only legal airframe). A what-if panel simulates gusts/fog/night to exercise the checks on demand.
• Simulated execution — the flight animates along the computed path; findings appear when the aircraft reaches them. Baseline findings are a deterministic seeded stand-in for an onboard CV model (labeled "sim" in the UI).
• Vision detection on real imagery — after launch, runway segments are fetched from Esri World Imagery (~0.3 m/px over KBOS), analyzed by Gemini vision with a bounding-box schema, and georeferenced to lat/lng. Vision findings appear on the map as labeled imagery strips alongside the simulated ones — same geolocate → work order → dispatch → resolve pipeline.
• Gemini copilot (optional, GEMINI_API_KEY) — plain-English mission requests are parsed into a structured plan via the Interactions API with a JSON schema; a keyword parser covers the same cases when no key is set. Completed missions get an LLM-written after-action report, or a template fallback offline.

Stack

Next.js 16 (App Router) · React 19 · TypeScript · Tailwind 4 · Leaflet/React-Leaflet (Esri World Imagery + OSM labels) · aviationweather.gov Data API · Google Gemini Interactions API (optional) · Vitest.

Run it

npm install
cp env.example .env.local    # optional: add GEMINI_API_KEY
npm run dev                  # http://localhost:3000

Verify

npm run typecheck   # tsc --noEmit
npm run lint
npm test            # 18 unit tests: geodesy, coverage, battery, compliance, solar math, weather
npm run build

Tests verify real-world invariants: published runway lengths reproduce from threshold coordinates within 2%, swath math covers full runway width, battery splitting honors reserve, compliance checks return GO/NO-GO correctly, and solar times match USNO for KBOS within 3 minutes.

Demo script

1. Ops dashboard — live KBOS METAR (28014G25KT grounds the quads today), fleet readiness.
2. New mission — type "FOD sweep on 04R" in the copilot, or pick manually. Note the 3-pass pattern on the satellite map and which airframes are eligible and why.
3. Preflight — CAUTION on wind; try the "what-if: fog" toggle to force a NO-GO, then back to live.
4. Create mission → Launch — 24-second simulated flight; findings pop as detected.
5. After-action report — Gemini-written if a key is set, template otherwise.
6. Work orders — findings become dispatchable work orders with coordinates.
7. Fleet — flight hours and battery cycles consumed by the mission.

Data sources

• Runway geometry: OurAirports (public domain) — ourairports.com/data
• Weather: aviationweather.gov Data API (live METAR)
• Sunrise/sunset: NOAA/SunCalc algorithm, verified against USNO for KBOS
• Airframe specs: representative generic classes — swap real specs in src/lib/fleet-seed.ts
• Imagery: Esri World Imagery export API (Maxar/NAIP etc., 0.3 m/px class over BOS) — used for Gemini vision segment analysis
• Simulated: baseline detection findings ("sim" label), perimeter fence line (computed 150 m buffer — replace with surveyed AOA boundary in production), what-if weather presets

Honest limitations

Simulated findings are placeholders for an onboard CV model. Gemini vision analyzes real orthoimagery, but base-map imagery is historical, not live capture, and a general vision model is not a certified FOD classifier — real deployment needs an onboard camera feed + purpose-tuned model. Perimeter patrol follows a computed buffer, not the surveyed fence line. Compliance guidance is a planning aid — the remote PIC remains responsible for 14 CFR Part 107 adherence and any COA conditions.