Observability
The monitoring plane for Ana — every run it made, how good the ontology underneath it is, and whether the next model change breaks either.
Analytics agents fail differently than dashboards.
A wrong answer looks exactly like a right one. Every thread is judged about thirty minutes after it goes quiet — four heuristic detectors plus an LLM judge at temperature zero, merged into one verdict and filed as a cause, a symptom, an outcome or a strength.
- Thirteen signal types, four categories
- A severity and a recommendation on each
- Strengths recorded, not just failures
Replay every thread, turn by turn
Duration split into LLM and execution, then broken down by startup, queries and tools. Every signal carries a recommendation, and one click opens a fix thread where Ana works the problem — usually a missing table description, or a definition that has drifted from the data.
- Duration split into LLM and execution
- Broken down by startup, queries and tools
- Every signal pinned to the moment it fired
Who is asking, and what about
ACUs and activity per person, each member carrying a flagged rate — threads flagged over threads analysed. Describe a topic in plain language and Ana tags every thread against it, then backfills your history and trends the volume against the prior period.
- ACUs and activity per person
- A flagged rate on every member
- Topics you describe in plain language
Find the one that is misbehaving
The same view for every playbook, dashboard, data app, agent and connector: what it costs, how often it runs, and whether anyone is using it. A check catches a playbook still pinned to a model your org retired, and offers the successor as a one-click fix.
- Playbooks, dashboards, data apps, agents
- Dead files and reclaimable tokens
- Playbooks still pinned to a retired model
Your monitoring stack, extended to Ana.
Thirteen product-metric gauges over a rolling thirty-day window — runs, threads, playbooks, Slack, agents, signal totals and warn rate, with signal counts labelled by type and category. Scrape them, push them, or send the audit stream somewhere else entirely.
Prometheus
- Mode
- Scrape
- Target
- /api/v1/metrics
OTLP
- Mode
- Push
- Interval
- 1–15 min
Audit logs
- Mode
- S3 or OTLP
- Cadence
- Incremental
Secure by default. Compliant from day one.
Everything the monitoring plane records is subject to the same controls as the data underneath it — enforced by your warehouse, not by us.
SOC 2 Type II
Independently audited security controls, verified annually.
SSO / SAML
SAML and OIDC with SCIM role sync from your identity provider.
HIPAA and GDPR
Full compliance with both, and BAAs available.
Granular permissions
Row- and column-level RBAC. If you can't see it, Ana can't either.
Runs in your VPC
Your cloud, your datacentre, or fully air-gapped with no outbound route.