SafetyShield

01The problem

Testing an AI system for unsafe behavior normally takes hundreds or thousands of repeated prompt evaluations. That cost puts careful safety testing out of reach for smaller teams and projects.

02Approach

• 01Designed the platform architecture: a Bayesian early-stopping evaluation engine, a probe dataset, a guardrail kit, and an examiner dashboard.
• 02Chose Bayesian early stopping to detect risky behavior with a fraction of the prompt evaluations.
• 03Planned a guardrail add-on measured side by side against an unguarded model in a live dashboard.
• 04Explained the architecture to the four-member FYP team to build shared understanding of evals and safety measurement.

03Targets

• Detect risk with at least 70% to 80% fewer prompt evaluations than a full sweep.
• Agreement of 95% or higher between the full evaluation and the early-stopped one.
• Reduce unsafe outputs from roughly 60% to about 10% with the guardrail in place.

These are demo targets for the 2026-2027 build, not measured results.