Autonomous Operations Agent (MSc, in progress)

The problem

Predictive models tell you what will happen, but production systems still need a human in the loop to act. This project asks where that loop can shrink safely.

Architecture

Agent observes a stream of sensor and system events, consults a policy model, and dispatches actions through a constrained action interface — with full audit logging and fallback to human review for low-confidence decisions.

Key decisions

Hybrid architecture: ML.NET for fast in-process predictions, Python for experimentation with newer model families, message queue between them so each side can evolve independently.

Outcomes

• In progress.
• Early results show meaningful reduction in human-in-the-loop interventions for routine cases while preserving escalation paths for novel situations.

What I would do differently now

This is the work my MSc and longer-term research interests point at: AI integrated into software that decides and acts, not just predicts.