03 — Technology
Every choicehas a reason.
Gemini and Google ADK provide bounded interpretation and planning. Deterministic services own permissions, state, validation truth, idempotency, budgets and approval integrity.
T.01The stack
Eight layers, no decoration.
- 01
Intelligence
- Gemini 3.7 Flash
- Gemini 3.5 Flash
- Google ADK 2.7.1
- google-genai
- Vertex AI
Bounded structured reasoning, explanation, and read-only assistant behavior.
- 02
Application
- Python
- FastAPI
- Oracle JET 20.1.3
- Preact
- TypeScript
The orchestrator, agents, policy services, adapters and validation logic.
- 03
Cloud runtime
- Cloud Run services
- Cloud Run Jobs
- Artifact Registry
- Cloud Build
Independent service identities, autoscaling, and execution isolation for finite work.
- 04
State and messaging
- Firestore
- Pub/Sub
- Dead-letter topics
- Transactional idempotency
Durable workflow state and at-least-once delivery that stays safe under retry.
- 05
Data
- SQL Server
- Cloud SQL for PostgreSQL
- BigQuery
- Cloud Storage staging
The measured source, the analytics target, and the staging between them.
- 06
Security
- IAM
- OIDC
- Secret Manager
- Firebase Authentication
- Policy-as-code
Secrets stay references; identity and policy are evaluated outside the model.
- 07
Observability
- OpenTelemetry
- Cloud Trace
- Cloud Logging
- Cloud Monitoring
Traces, logs and metrics that make a claim checkable after the fact.
- 08
Delivery and QA
- Terraform
- Docker
- Make
- pytest
- Vitest
- Playwright
- axe
Repeatable infrastructure and the gates that keep the estate portable.
T.02Rationale
Why each part of the system is there.
- Gemini 3.7 / 3.5 Flash
Bounded structured reasoning, explanation, and read-only assistant behavior.
- Google ADK
Agent construction, tool orchestration, and framework compliance.
- Cloud Run services
Independent service identities, autoscaling, and containerized control-plane deployment.
- Cloud Run Jobs
Execution isolation for finite data-movement tasks and independent job evidence.
- Firestore
Durable workflow state, transactions, registry records, evidence pointers and idempotency claims.
- Pub/Sub
At-least-once stage delivery, push authentication, retries and dead-letter routing.
- BigQuery
Target analytics store, validation queries, cost estimates and bounded query execution.
- Cloud SQL
Managed PostgreSQL source for the measured live migration path.
- Terraform
Repeatable infrastructure for services, identities, topics, subscriptions, storage and data resources.
T.03Research
Realistic enough to be credible. Controlled enough to be falsifiable.
WideWorldImporters
SQL Server legacy estate for discovery, lineage, risk and validation
Official Microsoft sample, MIT
PostgreSQL retail
Second engine and live Cloud SQL execution path
Project-authored fixture
Oracle SQL corpus
Dialect incompatibility and stored-procedure cases
Project-authored, informed by public examples
Airflow-style DAGs
Schedules and upstream/downstream dependencies
Project-authored fixture
Fault injection
Row loss, drift, duplicates, broken lineage, PII attempts, malicious instructions
Project-authored ground truth
No confidential or production client data is used. Because the harness knows which dependency should exist, which risk was planted, which action should be blocked and which check should fail, evaluation is repeatable rather than anecdotal.
Next
The console
Application Guide