Most café dashboards answer "what did I sell?" Ryo answers the question an owner actually has on Monday morning: "what do I order, who do I tell what, and what should I promote this week — and did last week's calls actually work?"
It reads the week's tills, the weather, and the menu's own economics, then forms opinions — a short ranked list of concrete Moves — turns them into a one-click Action Pack (supplier order, staff cards, social), and scores itself next Monday against what the tills actually did. No busywork, no neutral dashboards, no theatre.
🍵 Speaks like a calm, slightly dry Japanese sous-chef who hates waste. Ryo (涼 cool / refreshing · 量 measure) — the two things a café runs on.
The Dashboard — honest backtest, next-day prep sheet, daypart forecast, and the charts.
The Agent — Monday briefing, the agent pipeline trace, and ranked moves.
- 🎯 Action, not insight — every screen ends in something you can do today.
- 🔁 Closed-loop or it's theatre — recommendations are scored against reality, and the model adjusts on its own misses, in the open.
- 🧾 One-click Action Pack — Monday's moves become a supplier order (CSV), staff cards, and the week's social in one tap.
- 🧮 Honest about error — the backtest shows the error (±9.4%), not a flattering accuracy. Nothing rounded in our favour.
- 🗣️ Opinionated — Ryo ranks and commits. No "here are 40 charts, you decide."
Monday, 08:00 ☕ the owner opens Ryo
│
▼
Ryo reads: 13 weeks of tills · ☀️ weather · 🧮 menu economics
│
▼
Forms opinions → 5 ranked MOVES
↑ Stock Up ◎ Promote ↓ Cut ◇ Experiment
(each: impact line · reasoning · evidence · confidence)
│
owner taps ✓ Action / ± Modify / ✕ Ignore ← HUMAN GATE
│
▼
"Execute this week" → 🧾 ACTION PACK
supplier order (CSV) · staff cards · social · impact sim
│
▼
……… the week happens …………………………
│
▼
Next Monday: 🔁 SCORECARD — predicted vs actual
accuracy scored · the model adjusts
| Screen | What it does | |
|---|---|---|
| 🍓 | Agent | The Monday briefing — Ryo's opinion + 5 ranked Moves, each with a one-line impact (revenue / waste / margin), expandable reasoning + evidence, and a confidence score. Tap Action / Modify / Ignore. |
| 📈 | Dashboard | Honest backtest band · next-day prep sheet + daypart forecast · prep-plan changeover cards · weekday-revenue bars · 🥶 weather-sensitivity S-curve · 12-week trading history · unit-economics scenarios · forward-risk model. |
| 🧾 | Action Pack | One click → supplier order (copy + CSV) · printable staff cards · Instagram + Stories copy · expected-impact simulation. |
| 🔁 | Scorecard | The closed loop — what Ryo predicted vs. what the tills did, per-move accuracy, and how the model adjusted. |
| 📸 | Critique | Photograph the pastry case → Ryo reads the layout against demand and says what's in the wrong place, with the velocity to prove it. |
Editorial dark. Warm charcoal, a high-contrast Fraunces serif display with an italic matcha accent, monospace uppercase micro-labels, and sage-green + periwinkle data accents. Calm, precise, high information density — a dashboard that looks like it costs money.
The Agent tab runs a real multi-agent pipeline server-side:
Data (rules) → Insight (Claude) → Action (Claude) → Memory (rules) → Persona
- Data assembles the tills, weather, and unit economics into a context bundle.
- Insight (Claude Opus 5) reads it and surfaces the sharpest demand / attach / waste / margin signals.
- Action (Claude Opus 5) turns those into ranked Moves + the staff cards and social copy — via tool-forced structured output, so the shape is guaranteed.
- Memory feeds last week's predictions-vs-actuals back in, and writes this week's moves to a store.
- Order quantities are costed by the deterministic engine, not the model — the numbers are real.
Every run shows its own trace and a Live agents / Fallback badge. If there's no
ANTHROPIC_API_KEY (or the venue wifi dies), it serves an identical-shape
deterministic fallback — the demo never breaks.
npm install
npm run dev # → http://localhost:3000For the live agent pipeline, add a key (server-side only, git-ignored):
echo "ANTHROPIC_API_KEY=sk-ant-..." > .env.localWithout one, Ryo runs the deterministic fallback — the full demo still works, on mock data, with no wifi dependency. 🛜💀→🚫
- ✅ Real: the agent pipeline (Insight + Action on Claude Opus 5, structured output, memory feedback), the order-quantity math (
velocity × safety − stock + lead buffer), unit-economics scenarios, closed-loop feedback with persistence, and every chart (hand-built SVG, zero chart libraries). - 🔧 Mocked, on purpose: the Square + weather feeds and the Critique's vision pass are convincingly simulated rather than live — and the pipeline ships with a deterministic fallback so a demo runs with or without a key. Wiring real Claude vision and a Square feed is the first
TODO; the UI and data contracts are already shaped for it.
⚠️ The café Fern & Whisk and all figures are fictional on purpose — a demo build, no claims about a real business.
src/
lib/ryo.ts # domain types · café data · order engine · mock/fallback
lib/agents/ # Data→Insight→Action→Memory→Persona (server-side, Claude Opus 5)
app/api/brief/ # runs the pipeline; falls back to deterministic data
components/
AppShell.tsx # nav + closed-loop state (persisted)
Briefing.tsx # the Monday briefing (Agent)
MoveCard.tsx # a ranked move + reasoning + feedback
Dashboard.tsx # backtest · prep sheet · charts · economics · risk
charts.tsx # BarChart · AreaLine · SawtoothLine (inline SVG)
ActionPack.tsx # order · staff cards · social · impact sim
Scorecard.tsx # predicted vs actual, closed loop
Critique.tsx # photo → merchandising verdict
app/ # Next.js App Router · globals (the design system)
Built fast, built to demo. 🍵