A flight Starlink prediction tool hit Hacker News with 276 points and 362 comments. The tool estimates whether your flight will have working satellite internet based on flight number, date, and aircraft assignment. It does not give binary answers. It surfaces likelihood scores when data is incomplete.
This is a useful case study for agent data sourcing. The problem space has no official API, partial rollouts, time-sensitive equipment assignments, and a user expectation of confidence intervals instead of yes/no flags. The plumbing exposes how agents handle probabilistic infrastructure predictions when ground truth arrives days late.
The Data Model: Three-Tier Lookup
The tool runs a cascading filter:
- Airline check: Does this carrier have Starlink at all? If not, return 0% immediately.
- Aircraft body check: Does this specific model (A320, E145, 737 MAX) have Starlink? Some airlines retrofit entire fleets of one body type before touching others.
- Tail number check: Does this individual aircraft (identified by registration number) have the hardware installed?
The first two tiers are static lookups. The third tier is dynamic and time-gated. Airlines publish tail assignments 2-5 days before departure. Before that window, the tool calculates a probability based on fleet penetration: if 40% of United's 737-900s have Starlink, you get a 40% confidence score.
This is not a guess. It is a structured estimate derived from known deployment state and assignment randomness.
Scraping Rollout Schedules Without an API
Airlines do not publish structured Starlink deployment data. The tool maintainers built a database by:
- Monitoring press releases and investor calls for rollout announcements
- Scraping airline equipment pages for aircraft type assignments
- Cross-referencing passenger reports on forums and social media
- Tracking tail numbers from flight tracking APIs (FlightAware, FlightRadar24)
The refresh cadence matters. A weekly scrape misses mid-rollout changes. A daily scrape catches new tail assignments but burns API quota. The tool likely runs:
- Weekly: Airline and body-type policy updates
- Daily: Tail number assignments for flights in the 7-day booking window
- On-demand: User queries trigger real-time tail lookups if the flight is within 48 hours
This is the same pattern you see in hotel availability scrapers, concert ticket monitors, and GPU cloud capacity trackers. The data source is hostile (no official feed), the update frequency is unpredictable (airlines swap aircraft for maintenance or demand), and the user needs an answer now.
Handling the Cold-Start Problem
When an airline announces Starlink but has not published which aircraft get it first, the tool faces a cold-start gap. The database knows:
- Airline X has signed a Starlink contract
- Rollout starts Q2 2026
- No tail numbers confirmed yet
The tool can return one of three responses:
- Optimistic: "Airline has Starlink, but we don't know which planes yet. Assume 10% chance."
- Conservative: "No confirmed aircraft. Likelihood is 0% until we see tail data."
- Transparent: "Airline announced Starlink. Rollout in progress. Check back closer to departure."
The third option is correct. It avoids false precision and sets user expectations. An agent calling this API needs to know the difference between "no data yet" and "confirmed absence."
The API Contract for Probabilistic Predictions
If you are building an agent that books flights with connectivity requirements, the API needs to return:
{
"flight": "UA1234",
"date": "2026-11-20",
"starlink_probability": 0.65,
"confidence": "medium",
"data_freshness": "2026-11-16T08:00:00Z",
"tail_assigned": false,
"fallback_options": [
{
"flight": "UA5678",
"starlink_probability": 0.95,
"confidence": "high"
}
]
}
Key fields:
- starlink_probability: Float between 0 and 1. Not a boolean.
- confidence: Metadata about data quality. "High" means tail number confirmed. "Medium" means fleet-level estimate. "Low" means airline announced but no aircraft data.
- data_freshness: Timestamp of last scrape. Stale data is worse than no data.
- tail_assigned: Boolean flag. If false, probability is a statistical estimate. If true, it is a direct lookup.
- fallback_options: Alternative flights with higher confidence scores.
An agent can then apply a threshold policy: "Only book if starlink_probability > 0.8 and confidence is high" or "Accept medium confidence if departure is more than 5 days out and we can rebook."
Validation and Feedback Loops
The tool needs ground truth to improve predictions. Validation sources:
- Passenger reports: Reddit threads, Twitter posts, airline app reviews mentioning Starlink presence or absence
- Flight tracking APIs: Some services expose aircraft equipment changes post-flight
- Airline customer service: Agents sometimes confirm tail assignments if you ask directly
The feedback loop:
- User queries flight UA1234 on 2026-11-20
- Tool predicts 70% likelihood based on fleet data
- User flies and reports Starlink was present
- Tool logs (flight, date, tail, outcome) to validation database
- Next time that tail number appears, confidence increases to "high"
This is supervised learning with sparse labels. You cannot validate every prediction, so you prioritize high-traffic routes and recent flights.
Failure Modes
| Failure Mode | Cause | Mitigation |
|---|---|---|
| Tail swap 24 hours before flight | Maintenance, crew scheduling, demand surge | Refresh tail assignments daily in 48-hour window |
| Airline announces rollout but delays | Supply chain, certification, budget cuts | Mark airline as "rollout announced" with 0% until confirmed aircraft |
| Starlink hardware installed but not activated | Software issues, billing disputes, regulatory approval | Track "installed but inactive" state separately |
| User books based on 80% prediction, gets 0% | Statistical variance, bad luck | Surface confidence intervals in UI, offer rebooking policy |
| Scraper breaks when airline redesigns website | HTML structure change, anti-bot measures | Monitor scrape success rate, alert on parse failures, maintain fallback sources |
The worst failure is silent staleness. If the scraper breaks and the database stops updating, predictions decay but the API keeps returning stale probabilities. You need:
- Scrape health checks with alerting
- Data freshness timestamps in every response
- Circuit breaker logic: if last successful scrape is >7 days old, return error instead of stale data
Deployment Shape
A production version of this tool likely runs:
- PostgreSQL: Airline, aircraft body, and tail number tables with Starlink flags
- Redis: Cache for flight number lookups (TTL = 24 hours)
- Celery or Temporal: Scheduled scraping tasks with retry logic
- FastAPI or Flask: REST API for flight queries
- Sentry: Error tracking for scrape failures and parse exceptions
- Prometheus + Grafana: Metrics on query volume, cache hit rate, scrape latency
The critical path is the tail number lookup. If the flight is within 48 hours and the tail is assigned, you need a real-time lookup against the tail database. If the tail is not assigned, you fall back to fleet-level statistics.
Observability Hooks
You need visibility into:
- Scrape success rate: Percentage of scheduled scrapes that complete without errors
- Data coverage: Percentage of queried flights that have tail assignments
- Prediction accuracy: Percentage of validated flights where prediction matched outcome
- Query latency: P50, P95, P99 for API response times
- Cache hit rate: Percentage of queries served from Redis vs. database
A dashboard should show:
Scrape Health (Last 7 Days)
- United: 98% success, last run 2 hours ago
- Air France: 85% success, last run 4 hours ago (alert: parse failure)
- Qatar: 100% success, last run 1 hour ago
Prediction Accuracy (Last 30 Days, Validated Flights Only)
- High confidence (tail assigned): 94% accuracy
- Medium confidence (fleet estimate): 68% accuracy
- Low confidence (airline announced): 12% accuracy
Query Volume (Last 24 Hours)
- Total queries: 12,450
- Cache hits: 9,200 (74%)
- Database lookups: 3,250
- Tail assignments found: 1,800 (55%)
If scrape success drops below 90% for any airline, you need an alert. If prediction accuracy for high-confidence queries drops below 85%, your tail database is stale or incorrect.
When Agents Need Probabilistic Data Sources
This pattern applies beyond flight connectivity. Use it when:
- Ground truth is not available at query time (tail assignments, concert seat maps, GPU availability)
- Rollouts are gradual and uneven (feature flags, infrastructure upgrades, regulatory approvals)
- Users need to make decisions under uncertainty (book now vs. wait, pay premium vs. accept risk)
- Feedback loops are sparse but valuable (passenger reports, user confirmations, post-hoc validation)
Do not use it when:
- You can get authoritative data in real time (payment card BIN lookups, DNS records, SSL certificate validity)
- Binary answers are required (security clearances, compliance checks, legal eligibility)
- Probabilistic outputs confuse users more than they help (medical diagnoses, safety-critical systems)
Technical Verdict
Use this pattern when:
- You are building agents that need to reason about infrastructure state before official data is available
- Your users can tolerate probabilistic answers and understand confidence intervals
- You have a validation feedback loop to improve predictions over time
- The cost of waiting for certainty is higher than the cost of acting on a good estimate
Avoid this pattern when:
- You need deterministic answers for compliance or safety
- Your data sources are stable and authoritative
- Users expect binary yes/no and will not engage with likelihood scores
- You cannot maintain scraping infrastructure or handle frequent schema changes
The Starlink prediction tool is not about predicting the future. It is about surfacing what is knowable now, quantifying uncertainty, and giving agents enough structure to make defensible decisions under incomplete information.
Source Links
- Hacker News Discussion (276 points, 362 comments)
Top comments (0)