IROHA Order

A GPT-5.6 waiter that preserves hospitality while recommending dishes and taking confirmed orders for small restaurants.

Comment

Inspiration

Ordinary QR menus reduce staff workload by removing conversation. But for a small restaurant, conversation is part of the product: where today’s fish came from, what pairs well with sake, whether a dish is enough for two people, and what the restaurant recommends. Small restaurants also face staff shortages, language barriers, and the repetitive work of explaining the same menu throughout the day. We wanted to reduce that workload without turning hospitality into a grid of buttons. IROHA Order began with a simple question: can AI preserve the warmth and personality of a good waiter while making restaurant operations more efficient? Our goal was not to replace staff. It was to let AI handle routine menu conversations and order intake, while people remain responsible for hospitality, safety, and the moments that require human judgment.

What it does

IROHA Order is a GPT-5.6-powered waiter for small restaurants. A guest opens the service by scanning the QR code for their table. GPT-5.6 welcomes them in Japanese or English and continues as a free-conversation waiter. The guest is not limited to a scripted flow. They can ask questions such as: “Where is this sashimi from?” “What goes well with sake?” “Is this too much for two people?” “What would you recommend if I do not like spicy food?” GPT-5.6 answers using the restaurant’s own AI context, including menu descriptions, prices, ingredients, origins, pairing suggestions, portion information, recommendations, and service style. When the guest places an order, GPT-5.6 calls a structured propose_order tool with menu IDs and quantities. The server validates the items and calculates every price from the menu master. The guest then sees a readback card and must tap Confirm before anything is recorded. After confirmation, the order appears immediately on a front-of-house monitor and a high-contrast kitchen display. Guests can also check their order history, call a staff member, switch languages, and request checkout at the existing register. AI replies can be spoken using one consistent OpenAI voice, creating a more natural experience while previewing a future real-time voice waiter. Allergy and dietary-safety questions are handled differently from normal recommendations. A server-side guardrail bypasses the model and directs the guest to human staff.

How we built it

The guest experience uses vanilla HTML, CSS, and JavaScript, with a Node.js and Express server managing AI requests, session state, menu validation, price calculation, order confirmation, events, and operational data. Every genuine guest message is sent to the OpenAI Responses API with GPT-5.6. There is no client-side rule-based conversation engine. Each turn includes the restaurant profile, available menu, descriptions, origins, pairing information, conversation history, confirmed orders, and the current hospitality stage. GPT-5.6 can respond with natural text, suggested conversation chips, or a structured propose_order tool call. The model never provides authoritative prices. The server resolves menu IDs against data/menu.json, rejects invalid or sold-out items, calculates totals, and creates the confirmation card. Confirmed orders and events are stored locally as JSON, so the complete demo works without Google Apps Script or an external database. A shared /api/state endpoint aggregates table totals, order status, staff calls, and checkout requests. The store and kitchen screens poll this endpoint every three seconds. OpenAI’s speech API generates the spoken replies. We fixed the system to one voice and added audio caching to improve consistency and responsiveness. We used Codex throughout the Build Week development cycle to inspect the specification and existing static UI prototype, replace the scripted conversation, implement the GPT-5.6 tool boundary, diagnose session and checkout issues, build the operational screens, and create regression tests.

Challenges we ran into

The hardest challenge was balancing natural conversation with reliable restaurant operations. A good waiter must be flexible, but an ordering system cannot allow the model to invent menu items, trust AI-generated prices, or silently commit an ambiguous request. We solved this with a strict boundary: free conversation → structured proposal → server price calculation → guest readback → Confirm Conversation continuity was another challenge. After confirming drinks, the AI still needs to ask about food preferences and continue recommending dishes naturally. We wanted the service to have direction without returning to a rigid script. We achieved this by giving GPT-5.6 compact session state while allowing the model to decide the wording and respond freely to interruptions and unexpected questions. We also needed to distinguish dislikes from allergies. Dislikes are conversational preferences that the AI can use for recommendations. Allergies and dietary restrictions require human confirmation. We implemented deterministic server-side detection, a visible staff handoff, and prompt-level safety rules. Finally, voice introduced consistency and latency challenges. Fixing the voice identity, using low reasoning effort for restaurant turns, and caching repeated audio made the experience feel faster and more coherent.

Accomplishments that we're proud of

We are proud that IROHA Order is not just a chatbot or a visual prototype. It is a working end-to-end restaurant flow. Guests can have unscripted conversations about the menu, receive restaurant-specific recommendations, place structured orders, review prices, and explicitly confirm them. Those orders then reach real operational screens for the store and kitchen. We built three connected experiences—guest, front of house, and kitchen—without requiring an external back-office service. The core order receiver works locally, while Google Sheets remains an optional future integration. We are also proud of the system’s authority boundaries. GPT-5.6 provides the conversation and interprets the guest’s intent, but the menu master controls available items and prices, the server controls persistence, and the guest controls final confirmation. Most importantly, the product demonstrates that restaurant automation does not have to remove hospitality. AI can reduce repetitive work while still expressing the restaurant’s own knowledge and personality.

What we learned

We learned that useful restaurant AI requires much more than a strong conversational model. It needs trusted restaurant context, explicit session state, structured tools, operational displays, and clear boundaries between AI suggestions and authoritative actions. Natural language is excellent for understanding guests, but deterministic application logic is still essential for prices, safety, and confirmed orders. We also learned that a restaurant’s personality can be represented as context. Sourcing notes, taste descriptions, pairing suggestions, portion guidance, recommendations, and service principles give GPT-5.6 enough information to sound like a waiter from that restaurant rather than a generic assistant. Finally, we learned that the best role for AI is not always full automation. In IROHA Order, the AI handles the conversations it is good at and deliberately hands the guest to staff when human responsibility matters.

What's next for IROHA Order

The next major step is full-duplex voice conversation with GPT Realtime, allowing guests to speak naturally, interrupt the waiter, and receive lower-latency responses while preserving the same structured confirmation boundary. For production use, we would replace local JSON storage with a managed database, add authentication and multi-restaurant support, and provide store-side controls for sold-out items, order changes, and cancellations. We also plan to support reservation-based greetings and personalized hospitality with appropriate guest consent. An optional conversation mirror could store anonymized service interactions for aggregate analysis. Restaurants could identify frequently asked questions, unclear menu descriptions, popular pairings, and missed recommendation opportunities, then improve their AI context over time. This creates a feedback loop in which the AI gradually expresses the restaurant’s style more accurately. Long term, IROHA Order could help independent restaurants preserve and teach their hospitality across languages and new staff—not by replacing the restaurant’s personality, but by making its knowledge available at every table.

Built With

Share this project:

Updates