AI Restaurant
Intelligence.

The ops team was staring at six dashboards and still guessing how much chicken to prep. Three MCP servers + an LLM agent later, eight weeks later, Chronos-2 + LightGBM cut their MAPE by 56% and $2.1M of food waste a year with it.

key result 56% MAPE reduction, $2.1M savings

Restaurant Intelligence Platform — exhibit a / restaurant-forecasting-platform

MCP-powered restaurant business intelligence platform

Industry: Hospitality
Timeline: 8 weeks
Key result: 56% MAPE reduction, $2.1M savings
Stack: MCP servers, OpenAI Agents SDK, Chronos-2 + TimesFM 3.0 + LightGBM, DAIL-SQL NL-to-SQL, Langfuse, Ragas, PostgreSQL + PostGIS, Python

An LLM agent (OpenAI Agents SDK) drives three MCP servers — one for demand, one for forecasting (Chronos-2 + TimesFM as the foundation-model baseline, LightGBM where it still wins), one for DAIL-SQL-style NL-to-SQL — replacing a legacy chatbot and a stack of spreadsheets. MAPE fell 56% and the chain saved $2.1M a year in food waste.

The system runs 50K+ forecasts a day, answers 94% of ad-hoc questions in plain English against 847 query patterns (validated on an internal Spider 2.0-style benchmark), and pulls in event and menu signals so the suggestion on Thursday night isn’t the same as the one on a regular Tuesday. Every agent step is traced in Langfuse and regression-tested with Ragas + Promptfoo on every prompt change.

AI Delivery Approach

Audit the data first — POS, inventory, events, menu and weather all landed in one warehouse with contracts between the layers. Half the battle.
MCP servers, single agent — Three MCP servers expose NL-to-SQL, demand forecasting and event-aware recommendations as tools. A single OpenAI Agents SDK orchestrator decides which to call. The protocol is MCP and the agent is the LLM. We keep that line clean.
Evaluate against real plans — Every model output compared to the week the GM actually ran. Chronos-2 / TimesFM as the foundation-model baseline that the bespoke LightGBM has to beat. If the suggestion would have lost money, it failed the test.
Quiet rollout — Shipped to one restaurant, then six, then the network. Feedback from managers flows straight back into the agents’ tool definitions.

The hard part

Restaurant data is cursed. POS mis-fires on Friday night, loyalty joins with a different customer ID, menu items change mid-week. The hard part wasn’t the forecasting model. It was getting the agents to notice when the upstream data was lying, and to say “I don’t know” instead of inventing a number.

How it works

MCP restaurant intelligence

flowchart TB
  subgraph sources [Six operational sources]
    POS[POS]
    Res[Reservations]
    Ev[Events]
    Wx[Weather]
    Del[Delivery]
    Inv[Inventory]
  end
  POS --> contracts[Layer contracts]
  Res --> contracts
  Ev --> contracts
  Wx --> contracts
  Del --> contracts
  Inv --> contracts
  contracts --> pg[(PostgreSQL + PostGIS)]
  pg --> honesty{Data honesty gate}
  honesty -->|POS Friday misfire| idk[I do not know]
  honesty -->|loyalty ID mismatch| idk
  honesty -->|mid-week menu change| idk
  honesty -->|ok| dict[Data dictionary]
  dict --> agent[OpenAI Agents SDK orchestrator]
  agent --> plan{Tool choice}
  subgraph tools [Three MCP servers]
    plan -->|question| sql[MCP: DAIL-SQL]
    plan -->|demand| fc[MCP: forecast]
    plan -->|prep| rec[MCP: recommender]
  end
  sql --> validate{Validate IR}
  validate -->|ok| sandbox[Read-only SQL sandbox]
  validate -->|unsafe| refuse[Refuse / I do not know]
  sandbox -.-> patterns[847 query patterns / Spider bench]
  fc --> chrono[Chronos-2 / TimesFM 3.0 baseline]
  chrono --> beat{Beat baseline?}
  beat -->|yes| lgbm[LightGBM winner]
  beat -->|no| keepfc[Keep Chronos-2 path]
  rec --> eventsig[Event + menu signals]
  sandbox --> ui[Operator UI]
  lgbm --> ui
  keepfc --> ui
  eventsig --> ui
  refuse --> ui
  idk --> ui
  agent --> obs[Langfuse traces]
  ui --> ragas[Ragas / Promptfoo vs GM plan]
  ragas -->|would lose money| fail[Fail the suggestion]
  ragas -->|holds| toolsfb[Tool-def feedback]
  fail --> toolsfb
  toolsfb --> agent
  ui -.-> volume[50K forecasts a day]
  toolsfb -.-> rollout[Quiet rollout 1 then 6 then network]

Six sources hit contracts, then a data-honesty gate can refuse a POS Friday misfire, loyalty mismatch, or mid-week menu change with I do not know. Three MCP servers: DAIL-SQL against 847 Spider-style patterns, Chronos-2 / TimesFM 3.0 as the baseline LightGBM has to beat, and a recommender on event and menu signals. Quiet rollout 1 then 6 then the network. Langfuse plus Ragas close the loop against the GM's real plan.

What changed

Once the agents were live, forecasting accuracy jumped 56% and the network’s annual food waste bill dropped by $2.1M. Managers now ask their questions in English and get numbers they trust within seconds, instead of a deck, a ticket or a two-day turnaround.