H.E.S.T.I.A. — an AI restaurant management SaaS that turns point-of-sale data into daily decisions

Business context

A restaurant juggles dozens of data sources daily: the management system (RMS), supplier invoices, staff schedules, recipes, sales reports. An owner asking a simple question — «how much ham did we use this week?» or «what should I order next week?» — ends up exporting to Excel and calculating by hand. The RMS stays the operational core, but it records; it doesn't answer.

H.E.S.T.I.A. attacks exactly this problem: it doesn't replace the RMS, it connects to it. It syncs sales, inventory, orders, and staff data, then answers in natural language, computes a 7-day forecast, generates a replenishment list from current stock + consumption, and takes inventory from a photo. Operated by FAINTECH SOLUTIONS SRL (CUI 48078411) — our own product, not a client project.

Tech stack & key decisions

The RMS stays the core; H.E.S.T.I.A. connects, it doesn't replace

Restaurants don't switch management systems for the sake of a new app. We built a sync layer that talks to Syrve, Nexus, Poster, NextUp, Loyverse, iiko, and WinMENTOR through their official APIs. Sales and stock come exclusively from the RMS — never estimated. Connection passwords are never stored in plaintext.

Natural-language chat on a canonical data layer

«How much ham did we use this week?» is not full-text search — it's a question about entities (dish, period, consumption). We built a canonical entity layer (dishes, categories, days, sales channels) over the synced data, and the language model translates the question into queries on those entities, with the source always visible in the answer.

Forecast and food cost are computed from recipes, not from stock out

Food cost is computed from recipe × dishes sold, as a percentage of revenue — never from stock leaving inventory, which is a dirty measurement. The 7-day forecast and replenishment list (with estimated cost and supplier) are built on the same foundation: current stock + real consumption.

Photo inventory: OCR + matching against the RMS catalog

The owner photographs the inventory list; optical recognition extracts the items and matches them against the management system's catalog; quantities are confirmed and the inventory lands directly in the RMS. It eliminates hours of manual typing per stocktake. A second flow — automatic NIR (goods receipt) from an invoice photo — is in development.

Notifications on the channel where the owner actually is: email, Telegram, WhatsApp

Operational alerts (stock shortages, food-cost threshold breaches, leave requests) land on Telegram or WhatsApp, not in a dashboard nobody opens. Where the RMS doesn't cover staff scheduling, H.E.S.T.I.A. takes over: who is on shift, hours worked, salaries, leave requests approved from the app.

Key takeaways

1. Don't replace the existing system — connect to it. A HoReCa product that demands an RMS migration loses the customer before the demo. A per-RMS adapter pays for itself with every additional integration.
2. AI in a management product must cite its source. Plausible answers destroy trust; answers with a visible source («computed from recipes») build it.
3. Food cost is computed from recipes and sales, not from stock out. Stock is a lagging measurement; recipe × dishes sold gives the real cost of served food.
4. The photo as input removes the worst manual work. Photo inventory + RMS catalog matching saves hours of typing per stocktake and increases how often it gets done.
5. Telegram/WhatsApp notifications beat a dashboard. An owner doesn't open a new app daily; they reply to a message. An alert on the channel already in use is an alert that solves the problem.

↗ Visit hestiaai.ro