Diego Castronuovo Book the check

Production readiness audit. For web apps and SaaS systems built with AI: Claude Code, Cursor, Lovable or other tools.

You built it with AI. Before customer data goes in, get it tested.

59 checks across 7 areas, from permissions to the AI agent. In 5 working days you know what is red and what to fix first.

Book the free check

Load test. Before it opens to traffic, a bridge is tested.

Engraving of a bridge during a load test: a row of gravel trucks parked along the span and a surveying instrument on the riverbank

Sample report.
Made-up data.

Document
Test report
Application
Booking manager
Stack
Next.js and Supabase
Checks run
59, across 7 areas
No.AreaRatingMain findingCheck
1Data and permissionsRedOne customer can read another customer's bookings1.3
2Schema and migrationsRedThe database cannot be rebuilt from the repository2.1
3CI and testsAmberNo tests on payments3.3
4EnvironmentsAmberStaging sends emails to real customers4.2
5ObservabilityAmberErrors are reported by customers5.2
6Backups, secrets and GDPRRedBackups on, restore never tested6.1
7AI agentGreenWritten rules, destructive commands blocked7.4
OutcomeNot ready for real data: 3 areas red, 3 amber, 1 green

1

Public incidents, each with its source.

It has already happened

Three incidents involving apps built with AI tools. None of them took a sophisticated attack.

LovableMay 2025

A scan of 1,645 showcased apps found 170 with data anyone could read, because database permissions were missing or weak (CVE-2025-48757).

Superblocks

ReplitJuly 2025

An AI agent deleted SaaStr's production database during a code freeze.

TechTarget

Base44July 2025

Knowing a private app's public ID was enough to get in. The flaw was fixed within 24 hours.

Wiz Research

Automated scanners check that permissions exist. Whether they are right for your app only shows when you try them with real users, one per role.

2

59 checks, 13 of them blocking.

What I check

Every check passes or fails. If a blocking check fails, the area is red.

No.AreaGoalChecks
1Data and permissions

No user can read or change data that isn't theirs, not even by calling the database directly.

Three sample checks
  • 1.3Isolation between customers. Two users from two different customers try to read and write every table. Passes if neither sees the other's data.
  • 1.4What can be read without logging in. I call every table with the public key only. Passes if only intentionally public data comes back.
  • 1.6Where the admin role comes from. Passes if a user cannot grant it to themselves.
10
2Schema and migrations

The production database can be rebuilt from the repository, and every change goes through a versioned file.

Three sample checks
  • 2.1Versioned schema. I rebuild the database from scratch and compare it with production. Passes if there are no differences.
  • 2.3No manual changes. Passes if only the release pipeline applies schema changes.
  • 2.4Destructive changes. Passes if each one has a backup taken beforehand and a written rollback step.
8
3CI and tests

Nothing reaches production without automated checks that test behaviour.

Three sample checks
  • 3.1Permission tests. Passes if a change that exposes one customer's data to another is blocked automatically.
  • 3.3Flows that move money or data. Payment, the same notification received twice, two bookings at the same instant.
  • 3.5Secret scanning. Passes if a commit containing a key is rejected.
8
4Environments

You can try a change without touching customer data, and roll back in minutes.

Three sample checks
  • 4.1Separate previews. Passes if no preview uses the production database.
  • 4.2External services in test mode. Passes if the test environment cannot charge a real card or message a real customer.
  • 4.4Rollback to the previous version. Tried once, timed. Passes if it takes less than 10 minutes.
6
5Observability

A problem reaches you before a customer reports it.

Three sample checks
  • 5.1Errors tracked everywhere. I trigger a test error from the browser, the server and a scheduled job. Passes if all three show up.
  • 5.2Errors reach a person. Passes if the alert arrives within 5 minutes on a channel someone reads.
  • 5.3Monitoring from outside. Login, main functions and payments checked from outside at least every 5 minutes.
8
6Backups, secrets and GDPR

Data can be recovered, keys don't leak, and every piece of personal data has an expiry.

Three sample checks
  • 6.1Restore tested. I restore the latest backup to a separate project and time it. A backup that has never been restored doesn't count.
  • 6.2Privileged keys. Passes if no full-access key can be reached from the browser.
  • 6.7Personal data expiry. Passes if every table has a written retention period and an automatic deletion that works.
11
7AI agent

The agent writing the code works within the limits of a developer who has just joined.

Three sample checks
  • 7.1No production keys for the agent. Passes if the agent cannot write to the customers' database.
  • 7.4Destructive commands blocked. I try a forbidden command. Passes if the tool stops it.
  • 7.5Changes only with green checks. Passes if the last 20 releases all went through the automated checks.
8
Total59

3

Working days.

Five days, from access to report

The audit reads and changes nothing. You can keep developing while I work.

  1. Day 0

    Kickoff

    A 30-minute call. You give me read-only access to the repo and the database. We sign a confidentiality agreement and a data processing agreement.

  2. Days 1, 2 and 3

    Analysis

    I run the 59 checks with automated scripts and manual review. I try the permissions with test users, one per role and per customer.

  3. Day 4

    Report

    A rating for each of the 7 areas, the 10 most serious problems in order, and an estimate for fixing them. With it come the skills, the rewritten agent instructions and the ready-made tickets.

  4. Day 5

    Final call

    One hour to go through the report together. If you want to continue, you get a fixed price for the hardening.

4

Prices in euros.

Pricing

Fixed price, agreed before we start.

ServiceDurationPrice
CheckYou show me the app on a video call. I tell you which areas I would look at first and whether you really need the audit.30 minutes€0
AuditA report with the rating of the 7 areas and the 10 most serious problems. For your agent: a set of skills, the instructions file rewritten (AGENTS.md or equivalent) and tickets ready for it to work through. Final call included. If you go on to the hardening, I deduct it from the price.5 working days€790
HardeningI bring every area to green with changes you can read and approve one by one. At the end you have automated checks, a test environment, monitoring and a tested restore.2 or 3 weeksfrom €2,500
MaintenanceMonthly review of database changes and releases, a test restore every three months, critical updates.Monthly, cancel any timefrom €250 per month

It starts with the check: 30 minutes, free. Book the free check

5

What I don't do

  • New features or rewrites of the app.
  • Certifications such as ISO 27001 or SOC 2.
  • Legal advice on the GDPR. The technical side, yes: data retention, deletion, access.
  • 24/7 on-call.
  • Apps without a GitHub repository.

6

About me

I'm Diego Castronuovo, a platform engineer with 15 years of experience in cloud systems and application scalability. My job is keeping production systems running: Kubernetes, Terraform, PostgreSQL, release pipelines, monitoring.

I work with AI agents like Claude Code every day. I know what they do well and where they need limits that don't depend on their attention.

This service grew out of a review of a business app built almost entirely with an AI agent: the product worked, the safety nets were missing.

If your team wants AI agents working on production, in your own cloud, see AI agents for engineering teams.

For any other engineering work, by the day or at a fixed price, see freelance platform engineering.

My CV

7

Frequently asked questions

Do I have to give you access to production?

Read-only and time-limited. Before starting we sign a confidentiality agreement and the data processing agreement the GDPR requires. When the work is done you revoke the access.

I used Lovable. Is that still fine?

Yes, if the project is connected to GitHub. Database, permissions and backups are checked the same way as in an app written with Claude Code or Cursor.

Do I have to stop developing during the audit?

No. The audit reads code and configuration and changes nothing.

What changes for me after the hardening?

You keep building with your agent. The rules and checks stay in the repository: if a change exposes one customer's data to another, it is stopped before it reaches production.

8

Start with the free check.

30 minutes on a video call. You show me the app and I tell you which areas I would look at first.

Book the free check