Inside SentinelSRE
A four-part engineering breakdown of the exact architecture, database schema, verification invariants, and cost-routing decisions behind SentinelSRE—from a Friday afternoon 500 KeyError to a human-approved draft PR.
Why We Built SentinelSRE: Turning a Stack Trace into a Verified Draft PR
Why we pivoted away from generic AI PR reviewers, how we solved the GitHub-vs-runtime-error gap, and why we split the system into Phase A (Clerk + R2 Presigned URLs + 6-table Neon Postgres) and Phase B (Live LangGraph investigation).
Read engineering deep-diveThe Parallel War Room: 4 Specialists, 3-Tier Memory, and Live SSE
How LangGraph Send() runs Log, Git MCP, Runbook/Episodic Memory, and Security agents concurrently—eliminating confirmation bias and streaming every finding live over authenticated Server-Sent Events.
Read engineering deep-diveEliminating Hallucinations: Zero-History L4 Verification & Postgres HITL Gates
Why multi-agent pipelines get tunnel vision, how Python ast.parse() + an isolated fresh-context verifier catches bad diffs, and how PostgresSaver + interrupt() pauses execution on /payments and /checkout routes.
Read engineering deep-diveCost-Aware Model Routing: Cutting Investigation Spend by 65% Without Losing Quality
Routing triage and parallel extraction to Haiku while reserving Sonnet for synthesis and verification—with every token, millisecond, and dollar audited in Postgres llm_traces.
Read engineering deep-diveLaunch Incident #992 (KeyError: 'discount_code' on POST /api/checkout) in the Split-Screen War Room.