Blog

Two conversations in one section. The first is about money and timelines: what things cost, how an estimate is built, what quietly inflates it. The second is engineering: how the systems you cannot switch off are actually built. No filler, and no “it depends” in place of an answer.

Vortex. Paper
Choices & process15 min read

Vibe-coded app broken after launch? What to check and what to fix first

In short: an app built with AI tools and no engineer rarely breaks in the interface. It breaks where nobody looks: a database anyone can read, secret keys shipped to the browser, payments that never become orders, no backups and no monitoring. Fix data first, then money, then backups and monitoring. A full rewrite is not always needed: most projects can be rescued in parts.

Choices & process15 min read

Supabase security checklist: check RLS, keys and storage in your own app

In short: in a Supabase app, the browser talks to the database directly, and Row Level Security (RLS) is the only thing deciding who sees which rows. Check four things in your own project: RLS is on for every exposed table, the policies limit users to their own data, no secret key ships to the browser, and storage buckets are not public by accident. The Security Advisor in the dashboard catches most of it. If a secret key leaked, fix the cause, then rotate.

Choices & process9 min read

Rescue or rewrite an AI-built app: how to decide, and what to fix in what order

In short: most AI-built apps are cheaper to rescue in parts than to rewrite. A rewrite is justified when the data model is incoherent, when you cannot get access to your own code and infrastructure, or when the platform cannot support where the business is going. Decide on audit findings, not frustration, and fix in order of damage: data, money, backups, monitoring, then code.

Choices & process11 min read

How to hire someone to fix your AI-built app without losing control of it

In short: before anyone touches your app, make sure the code, database, domain and payment accounts are in your name. Give each person the narrowest role that lets them work, never your own passwords. Start with a fixed-price diagnosis that delivers a written, ranked list of problems and a plan with prices. Be wary of proposals that promise a rewrite or a timeline before anyone has looked at the code.

Guide10 min read

Building a fintech product: the complete guide

In short: fintech differs from ordinary development not in its technology but in the cost of being wrong. There is no “we’ll fix it later” here — a wrong number in an account balance is not a bug, it is a claim. Below: what such a product is made of, what the regulator actually requires, where budgets get buried and what this really costs.

Cost & timelines6 min read

What e-commerce development actually costs

In short: a store with a catalogue of up to 500 products, payments, delivery and an admin panel costs $6,000–12,000 and launches in about 5 weeks. A store with 1C exchange, customer accounts and loyalty points costs $15,000–30,000; a chain with warehouses, wholesale and an app costs $40,000–80,000. Below: what makes up the figure, what inflates it quietly, and where cutting corners is a mistake.

Cost & timelines5 min read

What mobile app development actually costs

In short: an app for iOS and Android from a single codebase costs from $12,000 for an MVP to $150,000 for a complex product with real-time features and offline mode. A full app with accounts, payments and push costs $30,000–60,000 and takes about 18 weeks. The main driver of price is not the number of screens but the complexity of the logic and the integrations.

Cost & timelines5 min read

What a corporate website actually costs

In short: a corporate website with content you edit without a developer and lead forms starts at $4,000 and launches in about 4 weeks; a 15-page site with a blog and two languages costs $7,000–11,000, a premium site with motion and integrations $12,000–18,000. A landing page costs from $1,500; a customer account area from $8,000.

Cost & timelines5 min read

What a Telegram bot and mini app cost

In short: a bot with menus, answers to questions and lead capture costs $1,500–3,000 and takes about 2 weeks. A bot with a catalogue, payments and broadcasts costs $4,000–8,000; a platform with branches, AI and integrations costs $12,000–25,000. A Mini App with its own interface starts at $9,000, a Mini App store with a cart and payments at $14,000. It is the fastest way to start selling online when a full app is not yet justified.

Choices & process6 min read

CRM or ERP: which one your business needs

In short: CRM handles the money that has not arrived yet — customers and deals. ERP handles what is already inside the company: stock, purchasing, production, finance. Almost always, start with CRM.

Choices & process6 min read

How to choose a development vendor and not get burned

In short: look past the portfolio and the price at how they build an estimate and what the contract says about source code ownership. A portfolio can show someone else’s work and a price can be lowballed, but those two things are hard to fake.

Engineering13 min read

A test that has never failed has never tested anything

Short version: an agent fixes a bug, writes a test next to it, and reports that the tests pass. Nobody checks whether that test was ever red. Put the old code back and run it: if it still passes, it did not test the fix. I built a tool that does this mechanically, ran it over 200 merged pull requests written by coding agents, and verified the result by hand before quoting a number.

Engineering12 min read

Tests do not prove the absence of a bug. This does

Short version: a test suite is a search. It can say “found” and it cannot say “none”. For most code that is fine — but not for a payment protocol, an order state machine, or anywhere two tasks touch shared state: there the bug is a rare combination of steps. Model checking visits every reachable state and answers the question tests cannot answer at all.

Engineering11 min read

Isolation levels: what your database actually admits

Short version: an isolation level is not a performance setting, it is the list of outcomes your application is allowed to see. Between “read committed” and “serializable” sit anomalies with names and with a price: a balance driven below zero, a shift with no doctor on it, a document numbered twice. Reading in the docs that the database “supports snapshot isolation” is not enough — that is a property of executions, not of text.

Engineering12 min read

The master dies: what actually happens to your data

Short version: “we run a three-node cluster” describes a configuration, not a property. The property appears once somebody has checked that a network split does not make two nodes believe they are in charge at the same time, and that a write already acknowledged to a client does not disappear along with the failed leader. That is checked by simulation, not by pulling a plug.

Engineering9 min read

A test fails once in three hundred runs. That is not weather

Short version: a test that fails once in three hundred runs has already found a real race — you simply cannot repeat it. What is random is not the bug but the interleaving the operating system picked. Take that choice away from it, and the failure stops being an anecdote and becomes an address.

Engineering7 min read

A retried request must not charge the customer twice

In short: the network does not tell you whether a payment went through — it simply goes quiet. The customer taps the button again, the mobile app retries on its own, and without protection you charge twice. Here is how idempotency works at the level of tables and code, and where it usually gets broken.

Engineering8 min read

LCP from 6.4 to 2.2 seconds: finding what covers the first screen

In short: a method for tracking down render delay, shown on our own site. The text of the first screen arrived from the server in 10 milliseconds but painted later: it sat behind a full-screen layer that only loaded JavaScript could remove. Moving the display condition from JavaScript into CSS took LCP 6.4 → 2.2 s and the performance score 73 → 98. Below is the diagnosis step by step, including the hypothesis that measurement ruled out.

Engineering7 min read

Moving a live database without losing writes

In short: “we will do it overnight on Sunday” is a bet, not a plan. The working technique is different: for a while the system writes to both stores at once, history is backfilled in the background, the data is reconciled, and only then does reading switch over. Up to the last step you can always go back.