Smoky Pepper
An AI agent running internal operations for a real restaurant.
The problem
Staff at multiple locations were sending the owner scattered messages throughout the day about things they needed. Supplies running low, questions about prep, requests that would come in at random times or not at all — someone would forget to mention they were out of something until it was too late. The purchasing manager had to track all of it mentally and compile orders from memory.
There was also the operational coordination piece: reminders that needed to go out at specific times ("start the chicken at 2," "check the rice before dinner rush"), training information that needed to reach staff in their language, and process updates that otherwise required showing up in person.
The question: could an AI agent sit in the middle of this and make it organized?
What was built
An AI agent that handles the staff-facing operational layer. The core of it:
Async message coordination. Staff tell the agent what they need whenever they notice it. The agent files each request through an Inventory API into a database. Later, it pulls from the database, organizes by supplier, and sends one structured message to the owner at the right time. Instead of random interruptions all day, the owner gets a single organized update when it's time to order.
Scheduled reminders. Cron-based, owner-configurable, per staff member. "It's 2 PM, make more chicken." "It's 6 PM, check rice for dinner rush." The owner sets them. The agent sends them.
Communication bridging. When non-standard training info or process changes need to go out in Burmese, the agent handles the translation. Day-to-day operations run fine in English and Thai — translation is for the cases where it's needed, not the default.
The architecture: LLM as the communication layer, tool calls to a deterministic database system. The agent talks to people. The database tracks the data. The guardrails keep it from doing anything it shouldn't.
This is staff and management-facing only. There is no customer-facing agent.
The POS system (separate build)
Shane also built the restaurant's POS from scratch — React Native, Expo, Supabase — after evaluating off-the-shelf options that didn't fit. The third version of his own design held up: staff with no prior POS experience figured it out without training. It's the system in use now.
What this demonstrates
An AI system running in real conditions. Real staff, real queries, real operational consequences. The agent has been live and in use long enough to be unremarkable, which is the goal.
The harness pattern. Narrow tool access — file inventory requests, send reminders, translate when needed — and nothing else. Reliability comes from the boundaries.
Requirements-driven design. The agent handles what staff actually do, not what was technically interesting to build. That's the work that happens before code.