How we build
How it was built, brief by brief.
Our demos are built by an AI coding agent working from a written brief. Below are those briefs, shortened, and what came out. They are the specs we hand the agent, not a chat log. A senior engineer writes each brief and reviews the result.
Both demos are our own products with made-up data. See them on the Work section.
Demo
Booking + payments app
The brief: goal
A working, good-looking demo product: a mobile booking web app, which also opens inside Telegram as a Mini App, where a customer of a salon, clinic or tutor picks a service and a slot, "pays" (test mode) and gets a reminder. An admin view lists the bookings. It must run locally with one command and as a static web build, so it works in any phone browser and, with mock launch data, outside Telegram too.
The brief: steps
- 1Vite + React + TypeScript + Tailwind, a static web app for any phone browser. In-memory / localStorage backend, seeded services and slots, no paid services. The same build also runs inside Telegram through the official Mini Apps SDK.
- 2Screens: services → slot picker → confirm + test payment → success; "My bookings"; an /admin bookings table. English copy, mobile-first, a light theme by default that follows the chat colors inside Telegram.
- 3Vitest tests for slot availability and booking logic. The production build must be green.
- 4README with what it is, a "Demo · own product · built in N h" label, run steps and two mobile screenshots (390 px wide).
The brief: rules
- Create no external accounts, deploy nothing, spend no money.
- Claim no real customers anywhere.
Demo
AI Lead Agent
The brief: goal
The flagship demo for small and mid-size businesses. A web app where inbound leads (contact-form submissions and emails) arrive. An AI agent qualifies each one (fit score 0–100, budget, timeline and intent extracted, plus the reason), drafts a reply and moves it through a pipeline: New → Qualified → Replied → Won/Lost. A human approves or edits the draft before it is sent. It must look like a real B2B SaaS product: clean, desktop-first, fine on mobile.
The brief: steps
- 1Vite + React + TypeScript + Tailwind, a static SPA with no server. The AI layer sits behind one interface, LeadAnalyzer. Demo mode (default): a deterministic rule-based analyzer with a short simulated "thinking" delay, so it works offline with no key. Live mode: calls the Anthropic Messages API from the browser with a key the user pastes in, kept in memory only, with the JSON answer validated by a schema.
- 2About 12 realistic, made-up leads of mixed quality: a hot SaaS founder with budget, a tire-kicker, spam, an agency partner, an enterprise RFP, a wrong fit. Screens: inbox with score badges and filters, lead detail with an editable draft and "Approve & send", a kanban pipeline, a dashboard, and a "Simulate new lead" button. Everything persists to localStorage.
- 3Vitest tests: scoring and extraction on the seed leads (hot beats cold, spam is flagged), pipeline transitions, schema validation of a live-mode answer. Tests and strict type-check green.
- 4A headless browser run on the production build: open the inbox, open a hot lead, approve the draft, move it to Won, simulate a new lead, reload and check the state persists. 0 console errors, no horizontal overflow at 390 px. Screenshots of the four desktop screens and the mobile inbox.
- 5README with what it is, the "Demo · own product · built in N h" label, how live mode works and run steps.
The brief: rules
- Companies and people in the seed data are clearly made up, and the README says so.
- Never commit or log an API key. No paid API calls while building.