MasarSaudi Arabia is opening to tourism at speed, and the discovery tools travellers arrive with all rank the same way: by popularity. Masar is built on the opposite premise. AI understands the traveller, a deterministic engine plans and optimises the whole journey against a real budget, and a measurement layer proves what actually moved — visitors off the crowded icons, spend toward the independent operators who were invisible to every discovery app.
Working MVP · Runs offline with zero API keys · Same question in, same answer out · Every estimate badged where it appears
Language models are excellent writers and unreliable accountants. Masar splits the work along exactly that line — and then keeps going past the plan, into the journey itself.
The traveller types freely in any language — no forms, no filters. The AI extracts origin, group, dates, budget, pace and interests into a typed, validated brief, and anything it inferred rather than heard is badged "assumed" and editable before a single stop is built.
Give it an origin and it plans getting there — mode, real door-to-door time, fare and costed transfers — not just the days. Travel and stay come out first; the days are planned against what is left. One recent plan: SAR 1,615.02 against a SAR 2,700 brief.
Congestion is scored as a function of time, so icons are punished for being crowded now, not for being icons. "Jabal Al-Qarah moved from midday to 16:40 — it is at capacity at noon, and 16:40 is when the light is best. A pottery workshop fills the gap." Same day, same budget, cooler caves.
Providers are matched on interest fit — not fame, not ad spend. There is no paid placement because there is no mechanism to sell one. A provider's single lever is publishing quiet-hour capacity, and a quiet hour is exactly what the engine is hunting for.
Masar reads tone, pace and state at each touchpoint and adapts to the person rather than pushing every visitor through the same journey. It mostly stays quiet, speaking only when a change is worth the interruption — one suggestion, with the exact cost, time and crowd impact. "Keep my plan" is a proper button, and it will never offer to skip the café you are sitting in.
Places store facts, not descriptions, so the write-up is composed for the reader at the moment they ask — the same cave reads differently for a German history reader, a Malaysian family and a Korean photographer. The Arabic name leads so it matches the signage in front of you, prices and times are protected from RTL reordering, and when it cannot write well in a language it says so.
All three surfaces read from the same engine. That is why this is a platform and not an app — every place, provider and setting is filed under a destination ID from day one, so city two is configuration rather than a fork.
A real trip: how you get there, where you sleep, what you do each hour, and what all of it costs. Travellers are never charged — paid recommendations would break the only asset we cannot rebuild.
They see which traveller interests they match, their own demand curve, fillable capacity per day and visits routed to them. The best pitch to a café is the café next door, whose quiet Tuesday mornings stopped being quiet.
The authority dashboard is not a report about the planner — it is the same planner, run across a cohort of trips. At the default setting, crowding at the famous sites drops by about half and spending reaching independents rises from under two-thirds to nearly four-fifths.
Push the lever to maximum and crowding keeps falling — but demand stops spreading out and piles onto a smaller set of quiet places instead. That caution sits above the default setting in the dashboard, not in a footnote, and the trade-off is itself the intelligence an authority is buying.
Providers store facts, not marketing copy — ownership class, provenance, quiet-hour capacity, opening hours, price band, visit duration, accessibility, interest tags, Arabic name. Ownership decides who receives demand, so it is checked rather than taken on trust.
A traveller photographs a problem — a taxi dispute, a damaged facility, a service failure, an unsafe situation. The platform reads the image and context, classifies the issue and routes it to the right party. Designed and specified; not in the shipped build.
Currency, timezone, bounding box, climate and weekend go into the registry alongside the public landmark graph. Nothing in the engine is edited to add a destination — a smoke assertion fails the build if any engine module imports a destination's data file.
Self-serve intake with ownership attestation, bearer-token auth and a review gate. An unreviewed submission cannot receive a single visitor under any policy we ship, so flooding the platform with fake businesses buys nothing but a queue for a human reviewer.
Free text becomes a typed brief, the solver simulates the timeline against hard constraints and the budget, and the AI narrates the result from the decision trace with a citation on every claim.
The day in progress, with the on-trip assistant reading tone and pace, tracking spend against the day, and re-planning only when a change clears the materiality threshold and the interrupt budget.
A nine-metric catalogue with unit, direction and basis on each; k-anonymity enforced at the query layer; and scenario comparison so an authority can see the expected impact on movement, distance and local participation at each setting.
Push the steering lever to maximum and crowding keeps falling — but demand stops spreading out. It piles onto a smaller set of quiet places instead. So we claim it moves demand off the icons and toward independent operators; we do not claim it spreads demand evenly across a destination. That caution sits above the default setting in the dashboard and at the top of the roadmap, and the maximum preset describes itself as the limit of the mechanism rather than a recommended policy. The trade-off is the intelligence an authority is actually buying: not "send tourists elsewhere", but "here is the expected impact on movement, distance and local participation at each setting."
Phase 1 is the planner and needs zero partnerships — success is follow-through, whether people actually walk the route, not sign-ups. Phase 2 is 20–40 verified providers, independents first. What we bring on day one is a working platform that runs offline with no keys, a frozen and reproducible measurement baseline, and a provider intake with a review gate already built.
Saudi-based delivery · Arabic-first · Working MVP · Runs offline for evaluation