bbconmigo

I defined the AI behavior framework (the three response states, the seven conversational components, the escalation logic) and drove product decisions from research evidence. Pediatric professionals validated my escalation classification against clinical criteria. A fullstack developer implemented the behavior in RAG and deterministic logic. Final product direction was signed off by the founder and co-founder. User Interface, User Experience and Design System was also designed by me.

bbconmigo began as a pediatric information product for parents of babies 0-12 months in Mexico. The initial hypothesis was simple: if parents had reliable answers in one place, they would feel more prepared.

Research changed the direction. Parents already had information sources. The product needed to help them interpret uncertainty, understand what mattered, and decide what to do next without replacing their judgment. The product is live and early-stage, built with a co-founder, a fullstack developer, and pediatric professionals.

The problem was not access to information. It was the gap between information and action when a parent was tired, worried, and responsible for deciding alone.

Field research with 180 parents across Mexico City, Guadalajara, and Monterrey showed five recurring clusters: anxiety, lack of immediate guidance, fear of emergencies, emotional support needs, and confidence in decision-making.

The interview sample focused on first-time parents between 20 and 40 years old, with a participant mix of 60% women and 40% men. The guide used eight questions to understand anxiety triggers, information-seeking behavior, AI familiarity, emergency consultation costs, and subscription willingness.

The product vision became a system that helps parents move from concern to a clearer next decision. bbconmigo should not answer every question the same way. It should understand the parent state, identify what context matters, communicate limits clearly, and preserve human judgment.

Only the insights that changed product direction carried forward into the design. The strongest pattern was not a request for more medical content. It was a need for structured interpretation during emotionally loaded moments.

AI behavior mattered more than UI polish because the product value lived inside the response posture. The same interface could either calm the parent, create false confidence, or delay action depending on how the system behaved.

Evaluation signals

The architecture was conceptual before it was technical: input becomes context signals, context changes safety level, safety level selects behavior, and behavior shapes the response and next action.

Decision boundaries

The system should not move directly from concern to reassurance when context is weak. It first determines whether it has enough signal to guide, whether it needs to clarify, or whether the situation should move toward professional care.

I designed the three-tier escalation classification based on the risk patterns surfaced in research. Pediatric professionals reviewed and validated the classification against clinical criteria before launch.

The diagram below shows how the same parent input: ''My baby has been crying for 2 hours'', can produce different system responses depending on context and signals available at that moment.

The strongest decisions balanced trust formation, cognitive load, safety, and conversion timing. No single metric could optimize the product alone.

Each decision below was driven by evidence I surfaced from research and usage data. The founder and co-founder held final sign-off on direction.

Interaction design was organized around product capabilities, not screens. Each interface pattern existed to reduce effort, capture context, or preserve decision agency.

The product moved from reusable UI components to reusable behavior components. That shift made behavior design part of the product system, not an afterthought attached to the conversation layer.

Ambiguous input iteration

A recurring abandonment pattern appeared when parents could not describe what they were observing. Vague input led to weak interpretation, weak interpretation led to generic responses, and generic responses ended the session.

The redesign changed the behavior. When the system cannot interpret a concern with enough clarity, it presents close-match options. The parent selects the closest situation, and the system continues with enough context to guide. Abandonment from interpretation failure stopped being a recurring pattern.

Behavior conversation logic

Instead of writing individual responses, the conversation assembled from seven reusable components: validation, reassurance, uncertainty, escalation, fallback, follow-up, and handoff.

I designed the AI behavior: three response states, seven reusable conversational components, and explicit rules for what the system should never do. A fullstack developer translated that behavior into RAG retrieval and deterministic logic.

These are early-stage signals, not scale claims. The useful question was what each metric changed in the product, not whether the product had already proven long-term growth.

The first lesson was that an AI product is not defined by the presence of a model. It is defined by how the system behaves when the user is uncertain and the answer has consequences.

The second lesson was that trust is easier to claim than to design. In this product, trust came from boundaries, transparency, escalation logic, and the ability to ask for less while understanding more.

The third lesson was that behavior scales better than individual responses. A response solves one moment. A behavior pattern can support many situations without becoming rigid.

How this got handed to engineering a master Figma file with documented user flows, micro-interaction notes, system behavior, and edge cases, plus a general view of the design system behind it.

• Repeated decision moments convert into paid trust, not only registration?
• An escalation logic reduces false reassurance without increasing unnecessary panic?
• Parents understand the difference between guidance, monitoring, and professional care?
• Response quality should be measured through perceived clarity, confidence, and next-action completion?
• How to keep iterating the generative AI knowledge base through specialty-specific pediatric audits, stress tests against pediatricians and specialty hospitals, and continuous updates that help the LLM classify each consultation by level and escalate appropriately?

The next stage is not adding more AI or build a complex design system. It is validating whether the system can remain useful, safe, and trusted as more parents, concerns, and edge cases enter the product.