STEP / 01
Accept structured telemetry with server-side validation.
CASE STUDY / 02 · INFRASTRUCTURE INTELLIGENCE
GRIDPULSE / Evidence before inference.
An evidence-first infrastructure intelligence control plane that separates observed telemetry, validation, prediction and explanation. The core design keeps the evidence record distinct from optional AI-derived analysis.
Next.js App RouterReactTypeScriptPostgreSQLZodAzure MLMicrosoft FoundryVercel
01 / Problem statement
Why it exists.
Infrastructure reports can be fragmented, noisy or delayed. A useful control plane needs to represent what was observed separately from what a model predicts, and must not turn a missing or invalid model result into a fabricated operational answer.
02 / System path
Architecture at a glance.
- NODE / 01Human telemetry
- NODE / 02Next.js API + Zod
- NODE / 03PostgreSQL evidence
- NODE / 04Azure ML
- NODE / 05Foundry explanation
STEP / 02
Normalize reports into a PostgreSQL-backed evidence layer and correlate reports within a bounded context.
STEP / 03
When configured, send structured features to Azure Machine Learning for a bounded score.
STEP / 04
When configured, use Microsoft Foundry / Azure OpenAI to produce an explanation grounded in supplied evidence and prediction.
03 / Engineering decisions
Trade-offs are part of the system.
DECISION / 01
Observed is not predicted
A model prediction is not proof that a disruption occurred. The validation layer and prediction layer are deliberately separate.
DECISION / 02
Fail closed on missing intelligence
If required optional AI services are missing, unavailable or return invalid output, the intelligence path should withhold the result instead of inventing one.
DECISION / 03
Bounded external services
Model version, confidence, generation time, validation state and evidence are distinct fields rather than a single unexplained verdict.
04 / Failure modes & limits
Where it can fail.
- AI integrations depend on correct environment configuration and external service availability.
- Rate limits, invalid responses and upstream timeouts should produce an unavailable/withheld state, not a fabricated score.
- The portfolio telemetry panel is explicitly mocked and is not an operational GRIDPULSE feed.
05 / Source map
Where to inspect.
src/app/api/README-documented routes for dashboard data, health, telemetry ingestion and intelligence.
src/lib/Database access, deterministic demo helpers and intelligence-service integration.
db/schema.sqlDatabase schema reference; incremental changes are under db/migrations/.
ml/ and docs/Model training/scoring contracts and integration, deployment and validation playbooks.
The map identifies only paths confirmed by available project documentation. Follow the repository link for the full current tree.
06 / External evidence