Use Case
Autonomous Operations
Your AI copilot that learns from your best operators. Tetrapus watches how your team works, discovers patterns, and proposes automation rules — then Claude agents execute them autonomously with full audit trails.
How automation emerges from operations
Operators work normally
An HVAC operator adjusts the thermostat setpoint every time CO₂ rises above 1000 ppm. Tetrapus captures every action silently.
TAI discovers the pattern
After observing 47 occurrences with a 4.2-second average response time, TAI identifies this as a recurring reaction pattern with 94% confidence.
A draft automation appears
The system proposes: "When CO₂ > 1000 ppm, set setpoint to 22°C." It appears in the operator's review queue — not deployed automatically.
The operator reviews and approves
They can approve as-is, modify the threshold, or reject. Approval boosts confidence by 1.2x. Three rejections permanently exclude the pattern.
A Claude agent takes over
The approved rule is deployed as an autonomous agent with scoped permissions. It can only act on the specific entities and fields defined in the rule.
Every action is audited
Every automated action is logged in a hash-chained audit trail. The operator can revoke the rule at any time.
TAI Engine pipeline
Key capabilities
| Capability | What it does |
|---|---|
| Pattern Mining | PrefixSpan sequence mining discovers recurring operator workflows — reactions to thresholds, multi-step sequences, and cross-entity cascades |
| Confidence Scoring | Every pattern is weighted by frequency, recency, and completion ratio. Confidence decays over time and responds to operator feedback. |
| Automation Synthesis | Reactions become alert rules. Sequences become state machines. Cascades become agent instructions. All generated as standard YAML configs. |
| Human-in-the-Loop | Every proposed automation goes through operator review. No automation is deployed without explicit approval. |
| Claude Agents | LLM-powered agents with scoped permissions. Each agent can only access the tools, entities, and fields explicitly granted to it. |
| Anomaly Detection | Z-score deviation and cross-domain Pearson correlation computed server-side in ClickHouse CTEs. |
| Audit Trail | Every automated action is logged in a hash-chained, tamper-evident audit trail. Full provenance from pattern discovery to execution. |
Claude agent framework
Tool Framework
JSON-schema defined tools for PaneAPI, ControlBus, ClickHouse, and config. Every tool call is captured by TAI.
Scoped Permissions
Entity-level restrictions: all, region, or group. Agents inherit the permission model of their principal.
Pane Introspection
Agents can read variables and call functions on any pane via PaneAPI — they see what operators see.
Policy Enforcement
Agent commands go through the same policy engine as operator commands. No bypass, no exceptions.
Ready to see the full picture?
One docker pull. Full stack in minutes. No build tools required.