Architecture decisions, made explicit
A useful architecture starts with a question, not a diagram. Explore the constraints, competing choices, and production evidence behind decisions that are expensive to get wrong.
Explore the library
Choose a problem.
Start with the challenge your team faces. Each resource examines the choices behind a possible solution, not just the technology in its diagram.
Research paper
Can an AI system act on company data without becoming an unchecked operator?
A useful assistant needs to retrieve information and sometimes call tools. The design has to separate what it may explain from what it may change.
You will examine: Identity, retrieval scope, tool permissions, approval, evaluation, and recovery.
Read the research paperResearch paper
When should a growing SaaS platform isolate a customer?
A shared product lowers operating cost, but one tenant can still affect another through data access, workload bursts, or recovery needs.
You will examine: Pooled versus isolated tenancy, workload boundaries, and evidence for moving between them.
Read the research paperWorked example · Illustrative
What must happen before a published schedule can safely change?
A closed field is a simple event with a complicated consequence: validity, human authority, concurrent edits, and communication.
You will examine: A narrow worked decision with a rule boundary, proposed workflow, and failure checks.
Examine the worked exampleResearch paper
How do you modernize a live platform without betting the business on a rewrite?
A replacement architecture is easy to draw. Transitioning users, data, and operations while the current system stays live is the hard part.
You will examine: Migration boundaries, staged releases, rollback, and the evidence needed to retire old paths.
Read the research paperWorked decision · Illustrative
Should an AI assistant investigate a production incident?
An incident commander needs a fast account of what changed, but plausible AI hypotheses can distract responders or trigger a dangerous action.
You will examine: Read-only evidence, source-backed hypotheses, explicit uncertainty, and the human command boundary.
Review the incident decisionWorked decision · Illustrative
When can a document update a business record?
Extraction can read an invoice or onboarding form, but ambiguous names, totals, and missing fields make an automatic write risky.
You will examine: Field-level evidence, deterministic checks, exception review, versioned writes, and audit.
Review the document decisionUse this in a review
Questions to bring to an architecture review.
Who owns the current fact?
A copied record is not necessarily the authority. Name the owner of the decision, its version, and how a conflict is resolved.
What is the simpler safe option?
Compare manual work, deterministic automation, and AI assistance before adding a model or another service.
Where does failure become visible?
Show stale writes, retries, partial delivery, and the signal that tells an operator to intervene.
What evidence would change the choice?
Set measurable conditions for scale, isolation, cost, and recovery. A reference design is a starting hypothesis, not production proof.
Facing a decision with no obvious right answer?
Share the current system, the constraint, and what failure would cost. We can help compare options and define what must be proven before a change goes live.