The Monday-morning plan chains get from an insight team.

A restaurant’s week is shaped by things outside its walls: a wet Wednesday, a match two streets away, a holiday, a rival’s new menu. Chains plan for them with data teams. An independent owner finds out on the night. Generic AI doesn’t help: ask it what to add this week and every restaurant gets the same answer, with no way to check any of it.

Every step runs on the model that suits it, on Nebius Token Factory. All model calls are measured and shown in the app’s engine room.

1. 1 · Context · Code, in parallel

   Locates the restaurant on OpenStreetMap, then fetches the 7-day forecast (Open-Meteo), public holidays (Nager.Date), events within 5 km (Ticketmaster) and every food & drink venue within 600 m (OpenStreetMap). Each item becomes a numbered source.

2. 2 · Menu photo · Nemotron 3.5 Super VL

   If the owner uploads a photo of the menu, the vision model transcribes dishes and prices into a structured list through a forced function call.

3. 3 · Research · Nemotron 3.5 Lightning + Tavily

   A fast agent loop (reasoning off) decides what to look up: the restaurant’s site and current menu, what guests and critics say, local news for the coming days, a same-cuisine rival’s prices. It searches, picks pages and reads them with Tavily Search and Extract; tools requested in one turn run in parallel.

4. 4 · Write · Nemotron 3 Ultra (Quick) or Nemotron 3 Super (Economy)

   The owner picks the writer: Ultra at medium reasoning effort, fast and the best average in my eval (Quick: about 1–2 minutes end to end, the writer itself ~30 s), or Super at medium effort, as faithful at a third to half the cost but slower (Economy: about 2–3 minutes; if it runs out of time, Ultra takes over). It writes the plan as one forced, streamed function call against a strict schema: demand moments, three dishes, rivals, an event, a post, a shopping list and a 7-day plan. Every claim must cite the numbered sentences of the sources it stands on.

5. 5 · Verify · Code, then Ultra once

   Code checks that every cited sentence exists, that every number in the claim appears in it, and that it says what the claim says (its specific words; numbers and names only when the page is written in another language). Claims about the restaurant must cite pages about it, an add-on goes with the dish it belongs to, and no date in the plan may fall before today. Failures go to Nemotron Ultra (low effort) once — one small call per claim, in parallel — with candidate sentences; a fix is kept only if it passes the same check. What still fails is shown as unverified.

• Code decides what is verified. No model is asked whether a claim is true; cited sentences and numbers are matched against the sources fetched in that run, and failures are shown, not hidden.
• Web pages are data, never instructions. Fetched text is passed to the models as quoted sources with an explicit rule to ignore instructions inside them, and the agent can only search and read — it cannot post, buy or send anything.
• No personal data. Tablecast reads public pages, open data and aggregate listings. It never collects information about individual guests.
• The owner decides. Dishes are bets to test for a week; nothing is ordered or posted without the owner.
• Credit where it’s due. Map data © OpenStreetMap contributors (ODbL), weather by Open-Meteo (CC BY 4.0), holidays by Nager.Date, events by Ticketmaster, web research by Tavily.

• Plans in English or Spanish, whatever language the sources are in.
• Works on a 390 px phone, and Save as PDF prints the whole plan.
• Everything works from the keyboard, with a visible focus ring; source chips open on keyboard focus as well as on hover or tap.
• With reduced motion turned on, nothing slides or drifts and scrolling jumps instead of gliding.
• axe-core finds no violations on the home page or on a full plan, on desktop and on a phone (checked 22 Oct 2026).

• Who pays: independent restaurants and small groups. A monthly subscription per location; the plan arrives every Monday.
• Why now: open models are good and cheap enough to run a full research-and-plan loop per restaurant per week for cents.
• Distribution: POS and reservation marketplaces, and food distributors — the shopping list is already an order.

Antonio González Fernández, Business Administration and Data Science student at CUNEF (Madrid). He previously built Supplora, an AI tool that reads supplier delivery notes for a family-run restaurant in Spain and loads them into its management system. Watching that restaurant plan its week on gut feel is where Tablecast started.