Chef Antonio

Problem

Voice assistants like Alexa and Google Assistant are everywhere but feel transactional, not human. We wanted to apply psychological principles so an agent could act as a warm coach, not a command line.

Goal

A flexible voice agent for cooking — focused on authentic Italian pasta — that adapts to different users and feels genuinely helpful.

Process

• Guidelines and user↔system perspectives
• User profiling (archetype, context, constraints)
• Agent personality definition (name, persona, coach behavior, traits)
• Conceptual model + initial dialogue tree
• Curated chef videos to validate recipe accuracy
• Iterative role-play of flows (dough, pesto, shopping list)
• Evaluation framework for user & agent responses

Challenges

Working remotely across time zones; managing many intents plus progress tracking; restart-on-error during testing — solved by merging all dialog trees so users can jump between options anywhere in the conversation.

Limitations

Accent variation; we deliberately avoided exact measurements (recommending options by serving size and equipment instead); and Dialogflow's limited predictive training-phrase capability.

Outcome

A coach-style voice agent that effectively helps novices learn three pasta skills, functioning as a buddy, coach, or assistant.