Your AI Architecture Co-Worker
Configured per workspace, on the model you choose. Operating in two directions — extracting candidate directives from your standards, and applying approved directives to every project.
The Two-Way AI Pipeline
Most teams reach for AI either at the front of governance (turn documents into rules) or at the back (review projects against rules). cajeX does both, on a single pipeline that traces back to the same approved directives.
Knowledge Base
Curate architectural standards, best practices, and reference materials
Directives
AI proposes candidate directives; architects review and approve
Review
Automated review of incoming projects against approved directives
Findings
Actionable findings with severity, confidence, and remediation steps
What the AI Does at Each Stage
The pipeline is not a black box. Here is exactly what happens when cajeX AI extracts a directive or reviews a project.
Knowledge base ingestion
When you add a document to the knowledge base, cajeX parses and indexes it so the AI can read it later. PDFs, Markdown, HTML, and plain text are all supported. Documents stay in the knowledge base whether or not extraction has run yet — they remain available as source material for future directive extraction and as supporting context during project reviews.
Directive extraction
When you run extraction on a knowledge base document, the AI ingests its full text in a single context window, identifies the rules it implies, and proposes them as candidate directives. Each candidate includes a suggested name, scope, severity, and a link back to the source passage. Architects review the candidates and approve, edit, or reject them before they enter the active rule set. The AI never adds a directive autonomously.
Project review
When a project is submitted for review, the AI loads your active approved directives and the project description. It evaluates each project claim against the directive set and surfaces violations as findings. Every finding references the specific directive ID it violates, so the audit trail stays clean and traceable.
Findings generation
Every finding includes a severity classification (informational, low, medium, high, critical), an AI confidence score, the directive it violates, and a suggested remediation step. Findings are mapped 1:1 to directive IDs so auditors can trace every recommendation back to an approved governance rule.
Watch an AI Review Run End-to-End
The same flow described above, on a real project. Submit a project, watch the AI evaluate it against the active directive set, and read findings as they appear with severity, confidence, and remediation.
Part 3 of the four-part Getting Started Guide. Watch the full walkthrough on /product.
AI Capabilities
Everything you need to run AI-powered architecture reviews at enterprise scale.
Knowledge Base Pipeline
Ingest standards documents, policies, and whitepapers in PDF, Markdown, HTML, or plain text. The AI parses the full document and proposes candidate directives, each linked back to the source passage.
Model Selection
Every workspace starts on GPT-5.6 Luna and can switch in Settings → AI to MiniMax M3, Kimi K2.6, Gemini 3.7 Flash, Claude Haiku 4.5, Claude Sonnet 5 and Claude Opus 5. Dearer models use the monthly budget faster, cheaper ones stretch it. Enterprise and Private Cloud workspaces can route AI calls through their own provider key instead.
Scoped Reviews
Run a review against your full approved directive set or a targeted subset. Scope by governance area, severity threshold, or project type — keeps reviews fast and tokens predictable.
Usage Dashboard
Monitor AI token usage, review counts, and per-workspace costs — broken down by model, task, and day. Set per-workspace quotas to manage your AI budget.
Example Findings the AI Generates
Real-shape examples of what an AI-generated finding looks like. Each one references the directive it violates and proposes a concrete remediation step.
Directive
“All public-facing services must implement circuit breakers with fallback responses”
Finding
Payment Gateway service has no circuit breaker on upstream calls to the fraud-check API.
Suggested fix
Wrap upstream HTTP calls with a circuit breaker library (Polly, resilience4j, etc.) and define a fallback response for fraud-check timeouts.
Directive
“PII must be encrypted at rest with per-tenant encryption keys”
Finding
User profile table uses a single application-level encryption key shared across all tenants.
Suggested fix
Migrate to per-tenant KMS keys with envelope encryption. Re-key the existing rows in a background migration.
Directive
“Service-to-service authentication must use signed JWTs with short TTL”
Finding
Order service authenticates to inventory service via a shared API key stored in an environment variable.
Suggested fix
Issue scoped JWTs from a service identity provider with TTL <5 minutes. Rotate the shared key out before deprecation.
Your Workspace, Your Model
GPT-5.6 Luna is the model a new workspace starts on. A workspace owner or admin can change it at any time in Settings → AI.
7 models to choose from
GPT-5.6 Luna, MiniMax M3, Kimi K2.6, Gemini 3.7 Flash, Claude Haiku 4.5, Claude Sonnet 5 and Claude Opus 5. The picker says which models use your budget faster and which stretch it.
Whole documents, not fragments
Long standards are read in full: extraction runs in parallel across the document, and a review covers every directive in scope rather than a sample.
Never used for training
Customer data, directives and AI output are never used to train or fine-tune any AI model. The security fact sheet names the provider behind each model.
An Assistant That Asks Before It Acts
Beside every screen, an assistant that knows cajeX and your workspace. A workspace owner or admin turns it on in Settings → AI.
It answers from your workspace
Ask how something in cajeX works and the answer cites the guide. Ask what is overdue, which frameworks suit you, or why a finding exists, and it answers from your own data.
It sees what you see
The assistant reads through your access. A confidential project you are not on is invisible to it too, and admin-only details stay with admins.
It proposes; you confirm
Moving a finding, rewording a draft directive, starting a review, editing a diagram: each is shown as a card first. Confirming runs the same audited action the buttons do, as you.
How it is built, and why it does not act alone: an assistant that asks before it acts.
AI Cost, in Your Hands
You choose the model, you see the usage, and nothing is spent beyond your plan unless you say so.
A monthly budget per seat
Paid plans run AI against a token budget that scales with seats, so a small review uses less than a large one. As a guide, a seat covers roughly this many reviews a month: 30 on Basic, 100 on Pro, 200 on Team and 280 on Enterprise.
Usage you can see
Settings → AI shows live usage for the month and the week, broken down by model, task, and day. The numbers are the ones cajeX itself uses to decide whether a job may run.
Overflow only if you approve it
When the budget is used up, a job already running finishes and new jobs wait for the next month, unless the workspace owner has set a paid-usage budget. Leave it blank and nothing is ever charged beyond the plan.
A monthly report
A consumption report for each month, as a PDF, shows what the workspace used and where.
AI Safety & Privacy
cajeX gives you full control over what data the AI sees, how long, and whether it sees anything at all.
- Per-workspace encrypted credentials. Where a workspace brings its own provider key (Enterprise and Private Cloud), the key is encrypted at rest with a per-workspace AES-GCM-256 key. No shared keys between workspaces.
- Disable AI entirely if needed. Workspaces with regulated data can turn off AI features completely. Every other governance module — knowledge base, directives registry, sessions, findings, reports — still works. See Regulated Industries for the BYOC story.
- Full audit trail. Every AI interaction — prompt, model, token count, response, timestamp — is logged for review. Auditors can trace every AI recommendation back to its full context.
Frequently Asked Questions
How does cajeX AI review architecture?
What AI model does cajeX use?
Is my data used to train AI models?
How accurate are AI-generated findings?
Can the AI extract directives from PDFs?
How does cajeX handle large standards libraries?
See cajeX AI in Your Stack
A 30-minute demo walks you through directive extraction, project review, and findings generation against your real standards.