Daedalus
Open

Architecture review that reads the design.

Daedalus reads the design document or drawing an application team sends, matches it to an approved blueprint, checks every control it is held to, and pins each finding to the component or flow it concerns. The review board decides on evidence it can see.

Open the workspaceSee a finished review
14approved blueprints to hold a design against
3frameworks scored on every review
82catalogued services, each on its own terms
Field Service PortalGoogle Cloud
Running0:00
0 of 21 components0 findings pinned
ReaderBlueprint MatcherPolicy FinderControl CheckerConsistency CheckerAdvisorReporter
Waiting for the file

One file in. A decision out.

A team sends one thing: the low-level design with its drawing inside, or the drawing on its own. Seven agents read it, hold it against the blueprint and the standards, and prepare the decision. A named person makes it.

  1. 01Read

    The Reader lifts the drawing from the page, names every component against the estate catalogue, and traces each flow, its protocol and whether it is encrypted.

  2. 02Match

    The Blueprint Matcher finds the closest approved blueprint, states its confidence, and lists where the design deviates and what each deviation risks.

  3. 03Check

    The Policy Finder retrieves the passages each control rests on. The Control Checker decides Pass or Fail with the evidence and its confidence, across six parameters.

  4. 04Trace

    Every finding is pinned to the component or flow it concerns. Where the document and the drawing disagree, the Consistency Checker says where.

  5. 05Decide

    The Advisor proposes fixes and the corrected architecture; the Reporter writes the summary. The chair approves, approves with conditions, or requests changes.

Reads what the team actually has.

Teams keep a design document with the drawing inside it, or they keep the drawing. They do not keep both. Daedalus takes whichever exists and says which one it read.

A design documentThe Reader finds the drawing on its page and lifts it out, then reads the prose around it. Every control cites the page and passage it rests on, and the document and the drawing are checked against each other.
A drawing on its ownThe review reads the components, the boundaries and the flows, and says plainly that no document was supplied. Nothing a document would have said is assumed.
Files are read in the workspace. The model is reached through the organisation’s own Foundry deployment and sees the design it is asked about, not the estate.
First page of field-service-portal-lld-v1.4.pdf
field-service-portal-lld-v1.4.pdf6 pages · drawing on page 2 · 26 sections
Architecture drawing warranty-claims-api-architecture.png
warranty-claims-api-architecture.png11 components · 11 flows · no document

Held to the standards the board names.

Security is scored against CIS Controls v8.1 and, for AI applications, the OWASP LLM Top 10. The other parameters are held to the provider's own framework — Google Cloud, Azure or AWS — with the estate's baselines on top. Every control shows the passage it rests on.

Pass 100 · Fail 0 · score = weighted mean by parameterA control passes or it fails. There is no third state to hide in, and nothing passes on silence.
SecuritySafeguards, policy alignment and the document against the diagram.
ReliabilityRedundancy, recovery objectives and dependence on others.
PerformanceStated targets, caching, scaling and where the work happens.
CostCapacity that runs when idle, retention, egress and cost visibility.
OperationsHow it is built, released, watched and recovered by people.
Estate fitServices approved for use here, used on their terms, built to the baseline.
Well-Architected reviewNorthwind, from the Google, Azure and AWS frameworks · 25 questions · reviewed quarterly
Decided by rule from the architecture model and the design document, so every review carries a reading for all five pillars, not security alone.Scored
CIS Controls v8.1Center for Internet Security · v8.1 · June 2024
Primary framework — every finding is recorded against a safeguard id, and every Northwind standard section cites the safeguards it supports.Active
NIST CSF 2.0National Institute of Standards and Technology · v2.0 · February 2024
Reached through Northwind's own CIS v8.1 → CSF 2.0 crosswalk. Unverified against the CIS-published v8.1 → CSF 2.0 mapping spreadsheet.Mapped
OWASP Top 10 for LLM Applications 2025OWASP Foundation · 2025 edition · November 2024
Applied automatically when a design contains a model, a retrieval index or an agent with tools; the AI Application Security Standard cites these items section by section.Applied to AI apps
NIST SP 800-53 Rev. 5National Institute of Standards and Technology · Revision 5 · September 2020, with Release 5.1.1 updates
Not mapped — add a mapping document to enable this view.Listed
ISO/IEC 27001:2022ISO/IEC JTC 1/SC 27 · Third edition · October 2022, with Amendment 1 (2024)
Not mapped — add a mapping document to enable this view.Listed
Google Cloud Architecture FrameworkGoogle Cloud · 2025 edition · continuously published
Its pillars are the source of the Well-Architected review questions, and its security pillar names the products the Advisor recommends for Google Cloud.Mapped
Azure Well-Architected FrameworkMicrosoft · Current edition · continuously published
Its pillars are the source of the Well-Architected review questions, and its security pillar names the products the Advisor recommends for Azure.Mapped
AWS Well-Architected FrameworkAmazon Web Services · Current edition · continuously published
Its pillars are the source of the Well-Architected review questions, and its security pillar names the products the Advisor recommends for AWS.Mapped

Every design is held against a blueprint the architecture office approved.

A blueprint is a reference architecture drawn in the same space as a review: the graph, the controls a design of that shape must satisfy, and the guidance the Advisor cites when a design deviates from it. One was assembled by Daedalus from reviewed designs that all failed the same way.

14blueprints in the library
3cloud providers
Customer-facing web application on Cloud Run
Web applicationGoogle Cloud · 10 controls · cited in 14 reviews
Internal GenAI assistant with retrievalderived
GenAI assistantGoogle Cloud · 17 controls · cited in 3 reviews
Event-driven data pipeline
Data pipelineGoogle Cloud · 7 controls · cited in 6 reviews
Internal API behind Apigee
API and integrationGoogle Cloud · 8 controls · cited in 9 reviews
Batch analytics workspace
Analytics and reportingGoogle Cloud · 6 controls · cited in 4 reviews
Containerised workload on GKE Autopilot
ContainersGoogle Cloud · 10 controls · cited in 6 reviews
Document intake with Document AI
Document processingGoogle Cloud · 13 controls · cited in 2 reviews
Hybrid integration with the on-premises ERP
API and integrationGoogle Cloud · 10 controls · cited in 11 reviews
Scheduled file-based integration
Data pipelineGoogle Cloud · 7 controls · cited in 9 reviews
Public content and media delivery
Content deliveryGoogle Cloud · 8 controls · cited in 7 reviews
Customer identity and access
IdentityGoogle Cloud · 10 controls · cited in 5 reviews
Agentic workflow with tool calling
GenAI assistantGoogle Cloud · 16 controls · cited in 4 reviews
Web application on Azure App Service
Web applicationMicrosoft Azure · 10 controls · cited in 6 reviews
Acquired workload on AWS
Acquired estateAmazon Web Services · 10 controls · cited in 9 reviews

Findings close on evidence, not on a promise.

After the decision, the application team answers each finding with what it changed: an applied plan, a policy export, a restore record. The checker reads the file, says which findings it closes, and refuses what it cannot use, with the reason.

Read, then matchedEach file is read and matched to the findings it answers. The line the checker relied on is kept with the record.
Refused with a reasonA drawing of the intended design, or a document of intent, is not evidence that a change is deployed. The checker says so, and says what would be.
Accepted gaps are signedA finding the team will not fix carries a justification, an owner, a date and the approver who accepted it.
Drawn as builtOnce the findings are answered, the architecture is redrawn: every evidenced change as built, every accepted gap marked, everything still open marked open.
The controls registerEvery control with its assessment, owner and citations. Exported as CSV for the delivery tracker.
The reportCover, summary, findings and fixes as an A4 PDF, drawn from the same record the screen shows.
The architecture after close-outThe updated drawing as a PNG the application team takes away, marked with what closed and what did not.

Seven agents, each with a job and a limit.

A run is seven agents, each with one job and a bounded set of things it may read. The chair sees what each one did, what it cited and what it found, as it happens.

01Reader

Turns the document and its drawing into one architecture model.

Names only what the catalogue knows. A service it cannot place is flagged, not guessed.
02Blueprint Matcher

Finds the closest approved blueprint and the deviations from it.

Reports its confidence. A weak match is shown as weak.
03Policy Finder

Retrieves and cites the passages each control rests on.

Quotes the passage. The citation stays with the control.
04Control Checker

Decides each control's status with evidence and confidence.

Pass or Fail, with evidence. Nothing passes on silence.
05Consistency Checker

Finds contradictions between the document and the diagram.

Raises only a contradiction it can point at in both the document and the drawing.
06Advisor

Recommends fixes and proposes the corrected architecture.

Proposes. It never changes a score or a decision.
07Reporter

Writes the review summary.

Summarises what the others found and adds nothing of its own.
Live calls fail loudlyWhen the model is called and the call fails, the failing step shows the real error and a retry. The review never quietly substitutes an answer.
Your endpoint, your tenancyThe model is reached through the organisation's own Foundry deployment. The key stays on the server and never reaches the browser.
A named person decidesApprovals, conditions and accepted gaps carry the approver's name, a justification and a date. The agents propose; the chair decides.