Sentinel
An AI assistant for incident investigation.
Concept · early developmentSentinel is an agentic AI project I am planning to help engineering teams investigate application failures. The first focus is payment failures and problems involving external payment providers.
The problem
When a payment or API call fails, engineers may need to search logs, trace a request across services, read provider errors, and work out whether the cause is inside their application or outside it. The evidence is often spread across different tools.
Sentinel aims to make that investigation more systematic and reduce the manual work needed to reach a useful explanation.
How the investigation is intended to work
The key idea is that an AI agent chooses its next check from the evidence it has found. This is a planned workflow, not a deployed system.
Describe the incident
An engineer provides a failed transaction, error, or incident description.
Collect evidence
Connected tools would gather relevant logs, API responses, error messages, and context.
Choose the next check
The AI agent would decide which investigation step to take based on the evidence found so far.
Find a likely cause
Sentinel would connect the findings and identify the most probable source of the failure.
Report the findings
The output would explain the likely cause, show supporting evidence, and suggest next steps.
Technical direction
The planned implementation starts with Python, connected investigation tools, and LLM APIs. The system would be modular so evidence collection, agent decisions, and reporting can be developed and tested separately.
Context retrieval, guardrails, and observability are part of the design work. Retrieval from past incidents may later use RAG or vector search where that is useful; those are options being explored, not completed features.
Where it stands
Sentinel is at the concept and early implementation-planning stage. The full agentic investigation system is not built or available to use yet.
Where it could go
The goal is an investigation assistant that can work with real application telemetry, help engineers understand failures, and produce findings they can verify.