Anu
Bilegdemberel

Problem

HANA already logs sessions, statements, and waits in detail. Operators still drowned in alerts, took too long to find root cause, and governance got shaky when people acted on a half-formed theory.

Solution

Python pipeline: extract features from SQL and SQLScript, combine rules with confidence-ranked root-cause guesses, spell out impact, and only run remediations from an allowlist with human approval and rollback. If we plug in an LLM, it polishes narrative text. It does not decide what is safe or what ranks first.

Evaluation & metrics

• Offline eval: precision@k on root cause, false-alert checks, zero tolerance on allowlist violations, narrative rubric completeness
• scripts/eval_run.py writes baseline vs improved profiles (e.g. alert dedup) to reports/eval-baseline-vs-v2.md

Security & reliability

No open-ended auto-fix. High-risk steps need a human. Decisions and actions append to an audit log. Secrets stay in .env or BTP destinations, not in git.

Problem

Every prompt or tool change nudged the agent off course. There was no steady signal before the build went to customers.

Solution

Multi-step LangGraph flows, golden-set regression tests, and LLM-as-judge checks in CI and staging, so each deploy had a clear pass/fail bar on the paths that matter.

Evaluation & metrics

• Core flows: regression detection went from days to hours
• Pass rates and latency budgets documented per workflow version

Security & reliability

Tool tokens scoped to least privilege, PII stripped from traces, rate limits and idempotency keys on anything that writes or calls out.

Problem

Plain vector search fetched answers that sounded right and were not. Docs changed between releases and did not always show up in retrieval. Access rules were not the same for every team.

Solution

Hybrid retrieval (dense plus keyword), metadata filters that match access policy, and a pipeline that refreshes embeddings on versioned chunks so answers stay grounded and citeable.

Evaluation & metrics

• Roughly 40% fewer off-topic or over-broad pulls vs a single-index baseline
• Grounded answer rate on a labeled eval set, checked each release

Security & reliability

Tenants separated at index and query layers, no raw PII inside embedding payloads, audit trail on what chunks were retrieved.