2026 · Full-Stack / Systems
IsMyTripSafe — Autonomous Destination Safety Intelligence
An agent-driven safety score for 1,000+ cities, built on government sources plus live web research.
2,000+ Monthly Users · Autonomous Pipeline
The problem
A traveller checking a destination faces fragmentary, inconsistent sources: the US, UK, and Canada each grade a country on their *own* risk scale, at their *own* cadence, in their *own* wording — and none of it says what the risk is like for someone actually walking around a specific city. "Is it safe to go there?" needs one answer with the reasons underneath.
The constraint
Research had to stay almost free. Reports run on OpenRouter's free-model tier, so each city must cost exactly ONE model call — every search and crawl happens in plain code (no model tool-loops), and results are cached with stale-while-revalidate so a returning visitor never refires the quota. And scoring had to be *non-compensatory*: a clean-air figure must never average away a "Do Not Travel" advisory.
What I built
A Next.js 16 plumbing where POST /api/research streams a report in ~10 s: geocode → fetch structured indicators (homicide, terrorism, road deaths, rule of law, corruption, air quality, nearby hospitals, live disaster alerts) → ingest US/UK/Canada advisory feeds → run a keyless web-research pass (Google News RSS + DuckDuckGo HTML + article extracts, one-per-domain) → one free model writes the verdict grounded in that dossier. The scoring engine is the differentiator: signals banded to 0–100, averaged *within* crime/conflict/advisory/institutions pillars, then combined across pillars with a severity-emphasizing power mean and non-compensatory hard caps (a level-4 advisory caps the score at 20). Big-metro inputs (Numbeo, FBI, city scale) let two cities in one country really differ.
The outcome
A free, always-current score + breakdown for 1,000+ cities that grew to 2,000+ monthly active users with zero human research after launch — the traffic story of the site, and the reason the whole scoring engine had to be correct enough to ship cold.