Toby Dobson

I built and operate a retrieval-augmented question-answering system over the official World Scholar's Cup syllabus, designed to answer only from source material, cite evidence, and refuse unsupported questions. I evaluated it against a 58-item test set for faithfulness, citation, refusal accuracy, ranking quality, and cost.

I also built YapYapGo, a live classroom speaking-practice SaaS with AI content generation, Stripe subscription lifecycle automation, signed webhook processing, OAuth 2.0 sign-in, and 170+ test files. I review AI output before it reaches classrooms and use testing and monitoring to catch regressions.

My seven years of assessment and academic coaching work shape how I evaluate AI systems: I look for fluent answers that are wrong, unsupported, or off-instruction. I treat generated output as a draft to test, document the evidence behind decisions, and work independently with production systems.