DataSentry™ is GSPANN's AI incident response system for data platforms. Its agents investigate each incident ticket, then either apply a governed fix or hand your engineers the evidence.
Your on-call engineers triage hundreds of data-platform alerts a day, and much of that volume is noise. When a real failure gets through, someone digs through siloed tools for hours while the business works from stale numbers. DataSentry™ runs that response as a governed workflow, from ticket to verified fix.
Both modes use the same agents and runbooks, and write to the same audit trail. Start as a copilot to build trust, and move toward autonomy as the results earn it.
Agents triage, investigate, apply a governed fix, and verify the outcome inside your policy limits. Engineers are pulled in only for what policy reserves for a person.
Agents do the investigation and bring back the root cause with a recommended next step. Your engineers review it and decide what happens.
Each ticket follows the same codified workflow, with guardrails and a human in the loop where policy requires it.
Normalizes the ticket and classifies it with a confidence score.
Specialist agents read job runs, notebooks, cloud logs, and table schemas to find the root cause.
Picks the next step from the evidence, such as a retry or a drafted code fix. When the evidence calls for a person, it escalates to L2.
Executes governed fixes and confirms the outcome before closing.
Posts the root cause analysis to the ticket and keeps it in sync. If the case still needs a person, it routes to L2.
An upstream column was renamed from event_timestamp to event_ts, but one downstream step was not updated. The Databricks job fails on every run, and the dashboards behind it go stale.
DataSentry classifies the ticket, finds the root cause, drafts the fix as a pull request, and waits for a person to merge it before verifying that the job succeeds. In the demo run shown on this page, the incident went from ticket to validated fix in 2 minutes 39 seconds.
Real screens from the DataSentry console. Click any screen to see it full size.
Every ticket that entered the system, what it was classified as, which team it went to, and where it ended up.
Specialist agents and their health. Each owns one job, such as classification, routing, Databricks investigation, remediation, or ticket updates.
The durable pipelines that execute each triage. See which step a run reached, and where it retried or stopped.
Live health of the whole system: tickets triaged, end-to-end triage time, token spend, and recent traces.
A scorecard for every agent, with its accuracy and 30-day trend.
See the active policy rules and every case waiting for a human decision.
A durable workflow engine coordinates specialist agents. They reach your systems through MCP connectors, and your chosen language models through a model-agnostic proxy.
| Layer | What It Does |
|---|---|
| Control Plane | A Temporal workflow engine runs each incident as a durable workflow, with human approval checkpoints where policy requires them. Its workflow UI lets you step through any run to debug it. |
| Agents Plane | Specialist agents, dispatched agent to agent (A2A) through the workflow engine. Examples include ticket intake, classification, Databricks investigation, and remediation. |
| LLM Proxy Layer | A model-agnostic gateway. Swap providers through configuration, with automatic failover when a provider hits a rate limit. |
| Platform Services | Agent registry, skill catalog, policy engine, and case chat. |
| Observability | Langfuse traces every prompt and its token cost. OpenTelemetry feeds latency and error-rate dashboards against your SLOs. |
| Integration Layer | MCP connectors and webhooks. The deployment shown connects ServiceNow, Databricks, a cloud platform, a GCP VM, GitHub, and a knowledge base. |
You can see where to trust automation and where to keep a person involved.
Reviewers confirm whether each agent was right, and accuracy is the exact-match rate against their verdicts.
An agent can propose a routing target without applying it, so you can score its decisions before you trust it to act.
Offline test scenarios replay fixed model replies through the real agent code. When one fails, the cause is in the code.
Policies are defined in one place and enforced at the workflow level. Every action is logged as it happens.
| Rule | Condition | Action |
|---|---|---|
| P1-DryRun | Priority is P1 or P2 | Force a dry run before execution |
| HITL-Gate | Remediation involves a job retry | Require human approval |
| Model-Override | Model confidence below 0.75 | Rules engine overrides the model decision |
| Token-Budget | Tokens per case above 50,000 | Halt and escalate to L2 |
| Transient-Auto | Runbook match and a transient error | Retry automatically, with no human needed |
Agents that inspect and classify. They cannot change anything.
Agents that suggest a change for a person or a policy to approve.
Agents that apply governed fixes, such as opening a pull request, inside the policy limits.
| Today | With DataSentry™ |
|---|---|
| Alert fatigue. Engineers triage hundreds of noisy alerts by hand every day. | Signal filtered automatically. Noise is suppressed before it reaches a person, and only real incidents escalate. |
| Hours to resolve. Investigation spans siloed tools, so MTTR is measured in hours. | Faster MTTR. Specialist agents investigate in parallel and resolve common incidents end to end. |
| Runbooks in wikis. Nobody follows them, so every response is improvised. | Runbooks enforced in the workflow. Each step follows the runbook, so every response to an incident class is the same. |
| Actions in chat. No audit trail, so compliance scrambles after every incident. | Full audit trail by default. Every action is logged, so the evidence is ready when compliance asks, with no extra work. |
Outcome figures are design targets and have not been measured in a production deployment.
Book a walkthrough and watch a failed pipeline go from ticket to verified fix.
