0-1 in 7 days

Overview

Mise consolidates POS, invoice, and recipe data to pinpoint losing dishes, explain causes, and offer fixes. I built it natively with AI from scratch, piloting in less than seven days, then rebuilt based on feedback.

The gap

Every restaurant margin tool shows what happened, but lacks guidance

I found lagging dashboards and outdated recipe calculators left restaurant operators reacting to food costs after the damage was done. AI-suggested price hikes do not draw on past data. Restaurant operators face tough questions alone, analyzing issues manually.

Existing platforms serve different roles: Toast IQ aids workflows, and FoodStorm handles ordering. Each excels in its area but lacks the full margin decision loop to decide, implement, measure, and learn. Mise fills this gap as an evidence layer beside the POS.

Built natively (with Claude)

I built Mise myself, directing architecture and decisions, with Claude writing the code.

A 7-day build required logic-first rules before screens. Then I worked like a PM with engineers: short prompts, review builds, redirects. The scope included problem framing, architecture, design system, product sections, and an AI agent. The real challenge was making decisions fast enough to keep the build moving.

Stress-tested on a synthetic persona

Before a real user, I built a synthetic restaurant GM with the skills and duties of running a high-volume restaurant.

Modeled on a high-volume revenue manager rather than a typical operator, the synthetic persona prioritizes margin dollars over food-cost percentages.

The persona, in raw form

Powered by a custom skill file, the AI enforced strict revenue logic like break-even checks and category rules. That is what let it push back on the product the way a real operator would.

An excerpt of the skill file behind the synthetic GM.

Approving a plan, in the pilot

Evidence-backed suggestions stage plans off-POS, deferring tracking until real-world changes are confirmed by a user.

Pilot outcomes

• Trust broke on the words, not the math. A Fried Chicken Half price recommendation cited a cost trend the data did not show. To avoid AI hallucination, I added server data validation and a strict 'cite source or refuse' guardrail.
• Visible evidence was positive. Shrimp & Grits showed clear approval as shrimp cost trends were visible and measurable. Evidence and trend lines became mandatory for recommendations.
• Silent failures cost more time than the ritual. Settings failed silently, demo recipes mixed with live data, and approving decisions could erase cards. I added save confirmations, strict live/demo separation, and reliable records. Gaining trust through guardrails was challenging.

The leap: tool to teammate

Moving on from rigid MVP directives, the post-MVP version simplified the interface to empower user decisions.

Pilot feedback confirmed core value but exposed cluttered views, weak iPad layouts, and manual review tasks.

Post-MVP, I moved the build to Perplexity Computer orchestration and Cursor, redesigning interaction to 70% human and 30% agentic flow. Plato handles prep and flags, shifting the GM from manual execution to strategic AI collaboration.

Overview

Before: I had dense cards, repeated counts, heavy borders.

After: I refined it to a single next action, with calm surfaces and source status in view.

Approving a plan

Before: I had a single-click approve dialog which needed more context.

After: I highlighted current vs proposed state comparison, weekly impact and evidence before approval.

iPad view with Plato

Before: I did not plan for iPad responsiveness in the MVP.

After: I accounted for Plato in the post-MVP with a dynamic content flow.

Plato demo

Product modules

It finds the risk before you do.

Mise automatically flags at-risk dishes against target margins, eliminating reports and spreadsheets entirely.

Two dishes at risk, the 68% target they are measured against, and the dishes themselves.

It shows its work

Every metric traces to raw costs, volume, and assumptions, letting operators trust data over blind faith.

Grilled Ribeye at 58.4% margin, 9.6 pts below target, and the open plan that follows from it.

It recommends. You decide.

Mise simulates margin and price impacts before execution, requiring recorded operator approval to go live.

The plan is modelled, and +$165.55 a week is visible before the operator approves it.

Plato, the AI agent

Plato had to feel like a teammate without ever acting like the boss.

Plato started as a line drawing beside an empty chat box. Moving past a generic chat box from the pilot phase, I gave Plato a memorable face that visually communicates its true operational state.

The character is my own mascot, a smiling plate, and it went through three versions before it worked in the product. I split the mascot into layered assets, using code to tie every expression directly to real-time system state.

Design system

The system had to move at the speed of the product without losing its source of truth.

During the MVP I built the Figma library, called Mise DLS 0.1.0. The post-MVP changed almost every surface, so the real question was how to update the system without breaking what the pilot relied on, and without Figma and code drifting apart.

I audited Figma gaps and built code-first foundations: semantic dark/light tokens and unified focus states.

Storybook became the working library. Each component page answers the same six questions: overview, anatomy, properties, states, accessibility and examples. Sourcing patterns directly from the production app kept design documentation grounded in shipped reality.

I imported v0.2.0 tokens into Figma without overwrites, archived legacy components, and published the library.

Mandatory WCAG 2.2 AA passes and code-first tokens established a single source of truth across design and code.

Design system repo

Results

A piloted loop, a rebuilt experience, and a product ready for its second restaurant.

Real-file edge cases in the pilot led to verification guardrails that automatically withhold unsupported claims. The post-MVP version applied these lessons, shipped with Mise design system 0.2.0, and went live on Vercel.

Reflection

Building natively with AI changed what being the designer actually meant, day to day.

I stopped making handoff artifacts and started making decisions, as each decision quickly became working software. My design judgment became the ultimate filter for setting firm rules, iterating screens, and rejecting weak assets. It also taught me to plan around tool limits when deciding where to plan and execute. A read-only Figma connector and no branching meant designing the release path as carefully as the screens.

The synthetic pilot proved the loop works. An industry run could bring in more feedback.

Live POS feeds and operator memory will test whether grounded briefs shorten review times and drive higher approvals. And expanding brand design could strengthen the product exposure.