BlueberrySoda

Blindspot is a monitoring agent for the problems that develop between sensors. Labs and server rooms may have only a few temperature and electrical meters, leaving equipment racks, desks, and power connectors unobserved. A normal reading at one sensor does not necessarily mean the whole room is healthy. Built with Meta’s Muse Spark, Blindspot investigates these gaps using five noisy sensor channels in a simulated lab. When a lightweight monitor detects an unusual pattern, it wakes the agent. Muse chooses which evidence to gather, checks whether sensors are trustworthy, investigates possible faults, and estimates conditions in hidden areas. Before proposing an intervention, the agent tests alternative actions in a digital twin. These include reducing noncritical load, capping rack power, switching on backup cooling, or requesting a human inspection. Every action request passes through a deterministic safety gate that checks the evidence, predicted improvement, and potential harm. The model cannot bypass this gate, and a human must approve an accepted request before it is applied. Follow-up monitoring allows further investigation if the problem continues. The application makes this process visible through an interactive room model, animated architecture diagram, live agent workflow, sensor graphs, human approval controls, and replayable decision records. Workspace accounts and role-based permissions separate operators, viewers, and administrators, while Neon Postgres stores completed workspace runs. Blindspot evaluates decisions against simulator ground truth rather than an AI judge. In a 20-scenario Muse evaluation subset, the agent detected 15 of 18 impending breaches early and avoided 12, with no alerts on the two no-fault controls. These are limited simulation results, not evidence of real-world safety performance. The project explores how an AI agent can turn sparse measurements into an evidence-backed investigation, while keeping physical calculations, safety constraints, and final approval outside the model’s control. Live runs use real Muse API calls; the room, sensors, and interventions are simulated. Blindspot is a decision-support research prototype, not a certified safety system.