Clients.Express

A Solo Marathon: From init to Production

March 25, 2026, first commit — a laconic “init”. 37 days later — a live B2B SaaS in production taking real payments. Between those two points: 373 commits, one developer and not a single framework.

Clients.Express is a platform for service businesses: barbershops, clinics, salons, auto shops, hotels, gyms. Under one roof:

• Client base — a full CRM with tags, notes and custom fields;
• Online booking — a calendar with five modes and conflict checking;
• Team and roles — owner, manager, employee, ownership transfer in one click;
• Public pages — a ready company landing page on its own subdomain;
• Billing — plans, client wallet, promo codes, payments via Monopay;
• An AI agent in Telegram — books an appointment from the phrase “book me with Olia tomorrow afternoon”.

Exactly what I build for clients as web apps and portals and AI agents — except this time the product is my own. Below: what it is made of — the numbers, the palette, the stack, the commit history, and the parts that broke against reality.

Why From Scratch, Not on Something Ready

For the first few days I honestly tried the “fast” route — Laravel + Filament. In a demo it is magic: one command and the admin panel is there. Then you need a non-standard billing flow, and you are no longer writing features — you are negotiating with someone else’s framework.

So the final stack is brutally simple: plain PHP 8.4 with no frameworks, Twig, MySQL over PDO, and a dozen hand-picked Composer libraries. My own router, my own admin, my own data model. No inherited technical debt — 62 core modules, and I can explain every one of them.

The code was written by two AI assistants — Codex and Claude Code — human-in-the-loop: the architecture, the module boundaries and the contracts stayed with me. AI accelerates the person who already knows what they are doing.

Three Features I Am Proud Of

The product has more than fifty features, but three of them are matters of principle.

• The AI agent in Telegram — not a menu of buttons but an agent loop with tool calls: the model checks free slots and creates the booking itself. No operation writes to the database without an inline confirmation from the user — a hallucination simply never gets the “Confirm” button.
• Billing with graceful degradation — the plan belongs to the company, not to the account. An overdue subscription blocks only new bookings for that company; the profile, the wallet and everything else keep working. No broken interfaces in front of a client.
• Grown-up security — three kinds of 2FA (TOTP, SMS, Passkeys via WebAuthn), geo-tracking of sessions, automatic logout on anomalies, and an active-session manager modelled on Apple’s: a freshly created session cannot terminate an older one.

Crash-Tested by Reality

For 37 days everything worked: demo accounts, a pretty grid of slots, tidy scenario tests. Then I visited the first live client — a barber — and heard: “listen, I can squeeze a client in between 14:15 and 14:40”. Between. Not at 14:00 and not at 14:30.

Not one of my booking patterns allowed for that — I was thinking in slots, and the barber thinks in living people. Two nights later barbers got manual editing of appointment boundaries in five-minute steps that does not break the price list.

The second thing to fall was the QR onboarding: the elegant “scan the code, land in the bot” flow died on a client with a single phone, who was shown that QR code on that same phone. A deep link instead of a QR — and the flow came back to life. UX is tested on people, not on the developer’s monitor.

This is the short version of the case. The full story of the marathon — LiqPay’s refusal, Monobank’s purring, and the barber who rewrote my architecture — is in the article “Solo Marathon: Building a SaaS with AI in 37 Days”. The product itself lives at clients.express — signing up is free.