Pipeline incidents are rarely solved by one error message. Engineers need run history, schema context, operating guidance and a clear line between diagnosis and execution.
Shivanand Kumar← Back to public work
Independent build · Synthetic data · CI passing
Data platform · Agentic AI
Data Platform Reliability Agent
A reproducible, deterministic-first reliability agent that turns synthetic events, contracts and runbooks into a bounded diagnosis; the optional LLM planner is not part of the current evaluation.
25/25offline regression cases passed
11tests passed
97%test coverage
Humanapproval boundary
Problem and ownership
A real engineering question, built to be inspected.
I designed and built the agent, API contracts, read-only tool layer, synthetic data, retrieval, safety controls, evaluation suite, tests and CI workflow.
Verified output
Concrete behavior, not a concept slide.
Abbreviated verified response from the schema-change test. The full response also includes direct evidence, citations, confidence, recommendations and the complete tool trace.
{
"severity": "high",
"issue_summary": "SCHEMA_MISMATCH",
"tools_used": [
"query_pipeline_events",
"inspect_pipeline_contract",
"search_runbooks"
],
"approval_required": true,
"execution_performed": false
}What the repository proves
Designed, tested and documented end to end.
- FastAPI and strict Pydantic request/response contracts
- Tenant-scoped SQL events and cross-tenant negative tests
- Contract comparison for missing and unexpected fields
- Qdrant runbook retrieval with evidence and citations
- Bounded planner rounds with deterministic final safety checks
- Approval-only remediation previews with no execution path
Architecture flowDesign decisions ↗
- 01
Validate the incident request and authorized tenant scope
- 02
Query synthetic pipeline events and inspect the expected contract
- 03
Retrieve the most relevant operational runbook
- 04
Synthesize evidence, likely causes, confidence and citations
- 05
Require human approval for any proposed state-changing action
Limits and next validation
What this project does not prove—yet.
- The 25-case suite is a deterministic offline application-regression gate. It does not evaluate the optional LLM planner, model prompts or tool-call quality.
- Required-tool checks confirm trace contents, and evidence-grounded checks confirm non-empty evidence and citations; they do not establish planner-selection quality, retrieval correctness or citation faithfulness.
- The default feature-hash embeddings are deterministic and reproducible, but they are not evidence of production-grade semantic retrieval quality.
- The demo accepts tenant scope in the request; a real service must derive authorization from verified identity and policy controls.
- The repository is a local, read-only demonstration using synthetic data. It does not execute remediation or claim production deployment.
Credibility boundary
Independent work, described honestly.
This is an independent portfolio implementation built from public career themes and entirely synthetic pipelines, tenants, events, contracts and runbooks. It is not employer production code or architecture.