EGPK.info — personalised aviation alert SaaS
NOTAM, weather and traffic ingest fanned out as per-pilot alerts so subscribers only see what affects their flying.
Pilots get drowned in raw NOTAMs and weather data. Most of it doesn't apply to their aircraft, currency or planned routes. The signal-to-noise ratio is so bad most pilots stop reading.
Built a pipeline that pulls source-of-truth data (NOTAMs, METARs, TAFs, ADS-B traffic), tags each item with the airfields/airspace/altitudes it touches, and matches against each subscriber's profile (home base, qualifications, intended flying window). Alerts go out via email or push only when the match is meaningful.
Light NLP classification on NOTAM text to disambiguate which airfield/airspace it actually refers to (NOTAM grammar is famously inconsistent).
Defined the matching rules with input from active instructors, and curated the ground-truth airspace polygon dataset.
Live SaaS with paying subscribers. Alert fatigue effectively eliminated for active subscribers — opens are high because the alerts almost always matter.
Initial ML classification was over-engineered — a 200-line rule engine with regex + lookup tables outperformed a fine-tuned small model on this domain. Removed the model entirely.
AI only earns its keep on the genuinely ambiguous 5%; the rest is rules + good data structures. Don't reach for ML when grep will do.
- 01Vertical SaaS wins by being radically narrow. 'Aviation alerts' is too broad; 'NOTAMs that match your aircraft and currency' is sellable.
- 02Per-user filtering is more valuable than smarter generation.
- 03Stripe + magic-link auth is enough plumbing for £5–50k ARR — don't over-engineer.