Offer · Describe a feature in the evening, test it in the morning
KaizoCRM 24
A CRM that grows with your company overnight. You describe a feature in plain language during the day, an AI coding model builds it at night, and local quality gates check it. In the morning you get a report and a preview, and nothing goes live until you approve it. Your company data never leaves your network.
For whom
Companies with 10–100 people and their own way of working, especially those that cannot send data to the cloud: law firms, accounting offices and manufacturers with know-how to protect.
Runs on
AI Office Box Pro or Pro+ for Cloud Coding; AI Office Box Max (96 GB) for Full Local. Compare AI Box packages →
System diagram
Inputs, outputs and interfaces
What goes into the AI Office Box, what comes out, and through which interface. Everything inside the dashed line runs in your network; the only outside connection is the coding model, at night.
Outside your network
Cloud coding model
Writes the code in a loop of code, tests and fixes. It receives only the anonymised specification, the module code and test results. In Full Local mode this connection does not exist: Qwen3-Coder-Next on an AI Office Box Max writes the code.
Your network: your data stays here
Inputs
Feature request
Your description in plain language, and your “I agree” on the acceptance criteria.
CRM web app in your LANCRM code and tests
The modules that may grow: reports, fields, views and automations. The core stays fixed.
Local Git repositoryDatabase structure
Tables and fields only, used to generate synthetic test data. Real records are never used for coding.
Read-only database connection
AI Office Box Pro / Pro+
- 1Specification: a local Bielik or Qwen3 model turns the request into acceptance criteria
- 2Preparation: synthetic test data, secrets scan, anonymised customer names
- 3Locked acceptance tests and the full regression suite
- 4Quality gates: lint, types, security scan, performance, independent reviewer model
- 5Preview environment, Deploy and one-click rollback
Outputs
Morning report at 7:00
What passed, what did not and why, with screenshots and a recording.
CRM web appPreview and Deploy
The new feature in the test environment; it goes live only when you click Deploy.
CRM web appNew feature in your CRM
Deployed to your live CRM, with one-click rollback.
Local deploymentEscalation to our team
A feature that fails two nights in a row goes to the KaizoBrain.ai team.
Remote service over VPN
Interfaces
| Interface | Direction | Protocol / format | What passes |
|---|---|---|---|
| CRM web app | You ↔ AI Box | HTTPS in your LAN, any browser | Feature requests, criteria approval, morning report, preview, Deploy and rollback |
| Code repository | AI Box ↔ CRM code | Git, local | Module code and tests; every change is a commit that can be rolled back |
| Database | CRM → AI Box | Read-only SQL connection | The schema, for synthetic test data; real records stay in the database |
| Coding model API | AI Box ↔ cloud | HTTPS through the firewall, to the model provider's domain only, 22:15–4:00 | Anonymised specification, module code, test results |
| Remote service | Our team ↔ AI Box | VPN, under the SLA | Updates, monitoring, escalated features |
How it works
Day, night, morning
- 1During the day
Day: you describe the feature
You write what you need in plain language, for example “a monthly report of customers with overdue payments, emailed on Monday”. A local language model (Bielik or Qwen3) turns it into a specification and acceptance criteria in plain language: “only customers past the due date”, “columns: customer, amount, days overdue”, “email on Monday at 8:00”. You click “I agree”, and the criteria become the contract the system is judged by.
- 222:00–6:00
Night: the feature is built and checked
At 22:00 the system prepares synthetic test data, scans for secrets and anonymises customer names. From 22:15 to 4:00 the firewall opens only to the coding model's provider: tests are written first and locked, then the model loops through code, tests and fixes. At 4:00 the connection closes, even if the work is unfinished, and local quality gates run.
- 37:00
Morning: you decide
You see a preview, screenshots and the list of criteria met, with a “Deploy” button and one-click rollback. If something did not pass, you get a clear reason, for example “the CRM has no payment due dates”. Code that fails the quality gates is never deployed.
Privacy
What stays with you and what goes to the cloud
| Goes to the cloud (night window only) | Never goes to the cloud |
|---|---|
| The feature specification and acceptance criteria, with customer names anonymised | Real data from your database: the coding model works on synthetic data with the same structure |
| The code of the module being extended | Passwords, API keys and logins (an automatic secrets scan runs before every upload) |
| Test results (which tests failed) | Company documents and correspondence |
Your data never leaves the company. Only code and the feature description go to the cloud, without your customers' data.
Quality gates
Five checks before anything reaches you
Acceptance tests
Generated from the criteria you approved (unit and end-to-end in Playwright) and locked: the coding model cannot change them.
Regression tests
All existing tests run again, so a new feature cannot break what already works.
Code quality
Lint, type checks, a security scan and a performance test.
Independent reviewer
A second model checks the code against the specification: does it do what you asked, and nothing more.
Evidence
Screenshots and a recording of the feature working in the test environment.
Night-time safeguards
Built to run while nobody is watching
Network window
Outside 22:15–4:00 the AI Box has no internet connection. During the day the system runs 100% locally.
Budget limit
A maximum model cost per night and per feature. Above it, the work stops and the report says so.
Attempt and time limit
For example 10 iterations or 90 minutes per feature, so one hard feature cannot block the others.
Watchdog
A supervising process restarts the agent after a failure and keeps the logs. If the model provider is down, the feature moves to the next night.
Fixed core
Only designated modules grow (reports, fields, views, automations). Core logic and the database schema change only with a person's approval.
Escalation to our team
A feature that fails two nights in a row goes to the KaizoBrain.ai team, as part of the subscription.
Two modes
Choose how the code is written
| Mode | Coding | Hardware | For whom |
|---|---|---|---|
| Cloud Coding (default) | A leading external coding model in the night window, the highest code quality | AI Office Box Pro / Pro+ | Most companies |
| Full Local | A local coding model (Qwen3-Coder-Next), lower quality, zero cloud connections | AI Office Box Max (96 GB) | Law firms, medical practices, companies that may not send code out |
“By day your system runs fully offline. At night it connects to the best models only to build the feature you asked for. In the morning you decide what goes live.”
All our offers run on the same on-premise platform: an AI Box in your local network with your own models. Your data does not leave the company, which matters for GDPR, trade secrets and medical data. We take care of the hardware, updates and secure remote service (VPN, SLA).
Start with KaizoCRM 24
Book a short discovery call. We measure your baseline KPIs first, so the result can be proven in numbers.
Request MVP