MLOps & AI Platform Engineer

Reproducibility

PROVEN, NOT ASSUMED

Destroyed cluster · images · build cache

make local-up followed by make local-test reached 11/11 acceptance checks in 901s (~15 min) after the cluster, images and build cache were destroyed first. Status is local-v1.0.0: the Kubernetes implementation (M0–M12) is validated and frozen. AWS (M13+) has not started — no cloud resource has been created, and Terraform is at the design/static-validation level only (fmt, validate, tflint; no plan, no apply).

Execution authority

SERVER-OWNED

LLM proposal → deterministic execution

Authentication, scope, policy, confirmation, revalidation, idempotency and business execution stay outside the model. The exercised D2c slice recorded 0 unauthorized mutations and 0 unsafe executions.

Root-cause judgment

DETERMINISTIC

Investigator proposes reads · deterministic RCA decides

A bounded, read-only investigator acquires evidence one legal read at a time; deterministic normalization, hypothesis rebuilding, verification and root-cause selection own the diagnosis. On the frozen blind ITBench-Lite TEST25 holdout it reached 21/25 (84%) exact-root agreement with 0 model calls. This is a measured 25-scenario benchmark result, not a universal production accuracy guarantee.

Multilingual reranker benchmark

MRR 0.367 → 0.956

220 questions · frozen multilingual benchmark

The multilingual reranker improved Cross MRR from the prior English reranker's 0.367 to 0.956 and reached Recall@5 of 1.000 on both TR-to-EN and EN-to-TR slices. These are retrieval/ranking metrics, not final-answer accuracy.

Routing control loop

DESIRED ↔ OBSERVED

Durable database state → router reconciliation

The database owns desired traffic; the router is restart-losable observed state. A worker-triggered reconcile tick repairs drift after a restart or failed push.

Execution authority

SERVER-OWNED

Model proposes SQL · deterministic admission decides

An ANSWER + SQL submission crosses parse, global policy, request-scoped relation authority, grain safety, EXPLAIN and a cost gate, and only an accepted immutable QueryPlan reaches the restricted read-only executor. There is no raw unsafe fallback, and the executor never runs SQL the safety service did not admit.

Sustainable isolated capacity

Workload-dependent

~1,050 evt/s at 42.8% fraud-eligible · ~1,600 at 0%

The processor has no single workload-independent ceiling, because fraud-eligible events take a longer path. The ~750 to ~1,050 evt/s (+40%) optimisation was measured on the 42.8% workload. Local isolated benchmark, not production capacity.