Venitech Talk to us Technology
We build most of the pieces behind the platform ourselves: the map, the menu pipeline, the AI layer, ranking, and how location travels. This page explains how.
Built on
Architecture
Restaurant OS, Veni App and the admin panel share one backend and one database. Each domain lives in its own module, in layers whose boundaries are checked automatically.
Every venue and group sits in its own data; every query runs inside a tenant boundary.
Orders, the kitchen, messages and notifications travel as event streams.
Terminals carry signed entitlements, so they keep working safely when the connection drops.
AI layer
An operator’s question is routed to the relevant tools among more than 70 across sales, stock, staff, menu, floor and finance. Tools that change anything do not run until the operator explicitly approves; input and output pass separate safety checks.
The primary model is Claude; if a model does not answer, the request moves to the next one in the chain. The choice of model is configuration, not code.
Dishes and venues are matched with multilingual embeddings, so different spellings of the same dish find each other.
Scout
A venue’s menu often sits behind a QR code, on one of dozens of different providers. Scout reads these menus with the community’s help and turns them into structured data.
A user inside the venue scans the table’s QR code with Veni. Location is verified within 100 metres.
The system recognises which provider hosts the menu. There are separate, tested adapters for 57 known providers.
Unstructured pages, menu photos and PDFs are read with Claude and turned into categories, items and prices.
A person reviews and approves every menu. An empty menu cannot be published.
With Claude
The vision and text steps of our menu pipeline, and Veni IQ’s primary model, run on Claude. The model turns a menu photo, a PDF or an untidy web page into categories, items and prices. The result is never published directly; it lands in a person’s review queue.
The model works from the menu itself rather than fixed rules, so menus from providers without a dedicated adapter can be read too. We also build our software with Claude.
Map
We generate the map ourselves from OpenStreetMap data and serve it from our own infrastructure. Venues, the moments people share and friends sit on the same map as separate layers.
The heritage layer comes from Wikidata. Every source’s licence is tracked separately, according to where that source may be used.
Location
Stays on the phone
40.9921, 29.0229
Sent to the server
8a1ec90213b7fff
On the fog map that clears as you explore, no coordinates leave the phone; only the ids of hexagonal cells of roughly 100 metres do. The day a cell was first visited is kept, the time and the route are not.
Live location is off by default. When on, friends see it snapped to a 50-metre grid, and it is deleted within 24 hours. In public mode no name, photo or venue is shown.
When finding friends from contacts, numbers are masked on the phone and the match is made on the phone (OPRF). The numbers of people who are not on Veni never reach the server.
Ranking
The discovery feed is ranked by a separate service we wrote in Rust. Proximity, liveliness and the quality of engagement are weighed together, with an exploration share so new venues also get seen. If the service does not answer, the app falls back to simpler ranking and the feed keeps going.
Resilience
Shared photos upload in the background and resume where they stopped if the connection drops. Terminals work offline and sync when the connection returns.
How we work
The contract between the app and the server is generated from a single schema. If a change would break an older app version, the build catches it.
Prices, thresholds and copy are not baked into code; they live as settings that can change.
Type checks, integration tests against a real database and code-quality scans run on every change. Code that computes money and permissions also has mutation and property tests.
The servers open no ports to the outside; traffic arrives through a Cloudflare tunnel. Backups run every night, every release passes a health check, and a failing one rolls itself back.