Build and defend a knowledge-driven agent that combines typed domain semantics, hybrid and graph retrieval, a bounded execution loop, and a reproducible harness.
Scenario
An operations assistant must answer service-impact questions, propose reversible remediation, and update reviewed incident facts. It must cite source evidence, respect ontology and policy constraints, survive dependency failures, and never commit an unapproved effect.
You will demonstrate
Project evidence
Each response is stored in this browser as you type. Include metrics, test output, or a decision rationale wherever the deliverable asks for it.
Define competency questions, stable identifiers, RDF or property-graph schema, provenance, temporal validity, and validation shapes.
Required evidence: Ontology, example graph, validation report, and three reviewed queries.
0/80 minimum characters
Implement lexical, dense, and graph retrieval with fixed query slices, authorization filters, citations, and an ablation.
Required evidence: Versioned corpus, query set, Recall@k and citation-support results.
0/80 minimum characters
Use typed state, explicit terminal reasons, budgets, idempotent tools, approval binding, checkpoints, and independent verification.
Required evidence: Transition contract plus success, exhaustion, denial, timeout, and resume traces.
0/80 minimum characters
Pin fixtures, isolate execution, inject failures, capture redacted traces, and support logical replay without repeating effects.
Required evidence: Harness policy, deterministic mocks, cleanup tests, and replay report.
0/80 minimum characters
Run normal, ambiguous, unanswerable, stale-graph, injection, and unsafe-action cases. Define quality gates, rollout stages, monitoring, and rollback.
Required evidence: Architecture diagram, evaluation report, threat model, and go/no-go decision.
0/80 minimum characters
Runnable Python lab
Approve only when retrieval, grounding, loop safety, and adversarial slices all meet policy.
Project defense
Rate the evidence, not your effort: 0 missing, 1 weak, 2 adequate, 3 strong. All criteria must be reviewed, but a low honest score does not get hidden.
Every generated claim retains source provenance.
Ontology, policy, and authorization remain separate controls.
The loop stops deterministically and verifies effects.
Retrieval and generation are evaluated separately.
Failure traces reproduce without secrets or repeated side effects.
The release decision names thresholds, uncertainty, rollback, and ownership.
Artifact
Submit a repository or archive containing the ontology, retrieval indexes or reproducible builders, typed runtime, harness fixtures, tests, redacted traces, evaluation report, and architecture decision record.
Anonymous mode stores only metadata and a fingerprint locally. Private upload requires an account and an explicit file selection. Uploaded notebooks are never executed.
Completion is controlled by stored evidence, deterministic tests, a decision defense, rubric review, and the artifact when required.
Optional cloud portfolio
An account is required. Submit only when the local completion gate passes. AI review is advisory and separate from deterministic completion.