All projects

Sentinel

An autonomous incident-response system where AI investigates and proposes, while deterministic Java guardrails retain authority.

Status
Live on Azure
Role
Designed and built the full system: control plane, data model, AI boundary, operator console, tests, and Azure delivery.
Distributed systems · AI safety2026
  1. 01Telemetry
  2. 02AI proposal
  3. 03Java guardrail
  4. 04Dry-run action
High-level flow only. Open the case study for decisions, evidence, and limitations.

On-call responders must connect alerts, deployments, metrics, logs, runbooks, and permissions under time pressure. Letting probabilistic output directly mutate infrastructure creates a second incident risk.

  • Separated AI proposal from execution authority with an in-process capability created only by a deterministic gate.
  • Designed at-least-once processing around Redis suppression, PostgreSQL uniqueness, RabbitMQ manual acknowledgements, and idempotent strategies.
  • Built a production-shaped digital twin and a public, dry-run operator experience.

The boundary

The model gets typed, bounded, read-only tools. It can classify evidence and propose a response, but it never receives a mutating tool. Every proposed action re-enters a deterministic guardrail gate.

Correct under retries

The system treats network and process failure as normal. Fast duplicate suppression lives in Redis; durable correctness lives in PostgreSQL; the queue acknowledges only after durable processing.

  • A database uniqueness constraint prevents duplicate incidents.
  • An action claim commits before an external side effect.
  • The execution strategy is idempotent across the crash uncertainty window.

Honest limitation

The public environment is a deterministic digital twin and stays in dry-run mode. It proves the decision path without granting anonymous visitors authority over real infrastructure.