Versatyl Group

01 / DataOne of the largest structured restaurant datasets anywhere

The Mealio Engine — a fleet of purpose-built ingestion pipelines — turns a messy spectrum of sources — menus, hours, happy hours, reservations, delivery, accessibility, photos — into clean, queryable structure before the algorithm ever touches it. The data engineering is the quiet win:

• 2B+ data points — structured fields, machine-applied tags, and dense embedding coordinates
• 1.9M menu items, structured and searchable down to the dish
• 20K+ venues deep in New York alone
• 100% neighborhood coverage across all five boroughs
• One of the largest aggregations of structured, non-review restaurant data in the world — compounding every day the engine runs

02 / UnderstandingFrom spoken thought to proven results

The engine understands structure, not keywords — and it shows its work:

• Full-sentence decomposition — one spoken thought, phrased however it comes out, understood as everything it asks for at once
• Language nuance survives the ranking — the subtleties that break keyword search don’t break Mealio: it knows a shrimp cocktail from a cocktail
• Evidence-bound answers — every match ships with the menu items, at the prices, that satisfy the ask
• Every algorithm change runs the Gauntlet — end-to-end suites of real cravings scored by an independent judge, plus adversarial red-team sweeps
• Algorithmic case law, kept in the Versatyl Playbook — every solved edge case becomes a permanent regression test; the engine never relearns an old lesson
• The whole stack is ours — no third-party search APIs at runtime, from understanding to corpus to ranking

No other consumer product converts open-ended food language into evidence-backed results at this breadth.

03 / LatencySpeed is a feature we refuse to trade

Depth usually costs speed. Engineering that trade away is the product:

• ~2s end to end — query understood, corpus ranked, receipts attached
• <150ms core ranking pass across 20K+ venues
• Proof in the same pass — menu evidence arrives with the results, not after them
• Live, computed results on every query — never a canned list

04 / CraftEditorial, engineered

A typographic design system where color belongs exclusively to computed results — if it's colored, the engine earned it. Answers surface the exact attributes you asked for. On-device dictation, native 3D, and a white canvas that lets the food do the talking. Designed like a publication, built like a search engine.

01 / The principleResponsibly, and with humanity in mind

While we believe AI can be a great tool to build new things and solve problems, we also believe that, like all technology, it should be used responsibly and with humanity in mind. We are committed to keeping our AI usage at 94% less than the next best search engines — Google Maps, Google Search, ChatGPT — while still being a better search engine than all of them for restaurants.

02 / The observationMost restaurant data does not need to be analyzed live

A large part of live search is re-analyzing the same data over and over: gathering webpages against the user’s search terms, re-reading reviews and publications, reconciling slightly different menu items, absorbing changes in SEO algorithms. The vast majority of the data that actually decides whether a restaurant matches your search only needs to be captured periodically — and once captured, it can be re-used for everyone else who searches.

03 / The designA database that already knows what it is for

Our search algorithm and database are built specifically for restaurants, so a language model never has to spend tokens working out what kind of thing you are asking about. The context is established by the structure itself. That is the leanest possible way to retrieve restaurant data — instead of a general search engine re-gathering, re-analyzing and re-generating a result set on every query.

04 / The measurementRoughly 17× less, and a better answer

Across approximately 15,000 test searches, producing a comparable pool of results with the relevant supporting information took roughly 17× the tokens for tools like ChatGPT and general search engines. The same design yields more accurate results, a wider breadth of results, and a pool not polluted by SEO or pay-to-rank. The restraint is not a trade-off against quality — it is where the quality comes from.