Connected operational model
Models services, deployments, changes, runbooks, and incidents as referenced Sanity documents instead of isolated text fragments.
AI engineering project
Incident Context
An incident-investigation agent that follows relationships across operational records, preserves the evidence behind every answer, and makes the boundary between confirmed facts and inference visible.
The problem
Similar symptoms do not prove the same root cause.
Incident evidence is spread across services, deployments, configuration changes, earlier incidents, and version-specific runbooks. Keyword matches can find related records, but they cannot safely establish causation.
Engineering decisions
Evidence trail
Shows the exact relationship path behind each investigation so operators can verify how the answer was reached.
Facts remain distinct
Renders confirmed evidence separately from model inference rather than collapsing both into authoritative-looking prose.
Gemini with structured output
Uses the Gemini SDK and validates answers, evidence, inferences, next steps, and sources through a Zod schema before rendering them.
Failure-aware dependency
Uses MCP Failure Lab to cover delays, hangs, malformed responses, connection loss, and live compatibility with the knowledge endpoint.
Production boundaries
Keeps credentials server-side and adds request limits, timeouts, explicit error states, and a health endpoint.
Investigation contract
Retrieve evidence. Trace relationships. State only what the sources support.
- 01
Retrieve
The agent queries a Sanity knowledge base of connected operational records through MCP.
- 02
Trace
It follows explicit references and records the source path behind each material claim.
- 03
Report
A validated response separates the answer, evidence trail, confirmed facts, inference, next step, and sources.