Numi

From personal experience to a validated problem

The initial idea was inspired by my own experience as a first-time mother. I encountered an overwhelming amount of contradictory sleep advice, while the products available to me mainly helped me record what had already happened. They did not help me understand why my baby’s sleep was changing or how I should respond. However, I did not treat my personal experience as evidence that the opportunity was broadly relevant. I used it as the starting hypothesis for discovery.

• A parent survey with approximately 30 participants
• Eight prototype and usability-testing sessions
• Competitive and market analysis
• Conversations with parents about their sleep-related decisions
• Input from baby sleep specialists
• Early acquisition and proposition testing
• Continued feedback from initial product users

This helped me distinguish between my personal frustration and the recurring needs shared by a wider group of parents.

What I learned

Parents were not primarily asking for more sleep data. They wanted help answering questions such as:

• Is this sleep pattern normal for my child’s age?
• Why has my baby suddenly started waking more often?
• Should I change the nap schedule or wait?
• What should I do tonight?
• How do I know whether a change is helping?
• When should I seek professional or medical support?

The strongest need was not tracking. It was interpretation, reassurance and decision support.

I therefore identified three recurring gaps in the existing experience:

Tracking without interpretation

Parents could record what happened, but they still had to determine what the patterns meant.

Advice without context

Articles and standard sleep schedules could not account for a child’s age, recent sleep, routines, environment or developmental changes.

Support without continuity

Professional support could be valuable, but it was not available continuously and was often financially inaccessible.

Core job to be done

Help parents move from recording sleep to understanding it, and from uncertainty to a practical next step.

Numi’s user personas

Based on insights from user interviews and surveys, I created three core personas representing Numi’s primary users. Because the product also offers in-app human support, I included a dedicated consultant persona to reflect the needs of baby sleep specialists using the platform.

Personalised sleep guidance that adapts with the child

Give every parent access to personalised, trustworthy sleep guidance that adapts as their child grows.

The product would combine sleep tracking, developmental context and conversational support in one continuous experience. Instead of simply showing parents charts, Numi would help them understand the patterns behind those charts and translate them into practical next steps.

Guidance over tracking

Every piece of data should contribute to an explanation, recommendation or decision. Tracking should have a clear benefit for the parent rather than becoming an administrative task.

Explain, do not command

The product should explain why it is making a suggestion rather than presenting advice as an instruction. Parents should remain in control of the final decision.

Personalise with restraint

The experience should use relevant context to make guidance more useful without creating an illusion of certainty. Personalisation should improve relevance, not make unsupported claims.

Support, do not judge

Many parents already feel pressure around their child’s sleep. The language should reduce anxiety and avoid framing normal sleep difficulties as parental failure.

Design for exhausted users

The experience needed to remain understandable when used at night, under stress or with limited attention. Logging, navigation and guidance had to be simple and immediately actionable.

AI with clear boundaries

The product should distinguish between educational sleep guidance and medical advice. It should acknowledge uncertainty and direct parents to professional support when appropriate.

From a broad family-support vision to a focused MVP

The original opportunity space was large — sleep prediction, feeding support, room-condition analysis, developmental milestones, human coaches and connected sensors. Building all of these at once would have made it difficult to validate whether the central proposition was valuable.

Record what happened → understand the pattern → receive contextual guidance → try an adjustment → observe the result.

Included in the first product

• Child profile and developmental context
• Sleep and nap logging
• Feeding logging
• Sleep-history overview
• Pattern-based insights
• Conversational AI support
• Contextual sleep guidance
• Developmental milestone content
• Basic recommendation and follow-up flows

Prioritised for later validation

• More proactive sleep predictions
• Advanced recommendation personalisation
• Environmental room assessment
• Human sleep-coach escalation
• Connected sensors and smart-home integrations
• Broader parenting and developmental support

The challenge was not adding a chatbot

A generic conversational interface would not be sufficient. Parents needed the AI to understand their child’s recent sleep, identify relevant patterns, ask for missing information and provide practical guidance — without sounding overly certain or clinical. It also needed to work for parents who did not know what question to ask.

How might we make AI guidance feel contextual and useful while preserving clarity, trust and parental agency?

1. Collect context

• Child’s age
• Recent naps and night sleep
• Wake times
• Feeding information
• Parent-reported routines
• Current concerns
• Developmental context

2. Identify patterns

• Changes in total sleep
• Long or short wake periods
• Shifting nap timing
• Repeated night waking
• Irregular bedtime patterns
• Recent developmental transitions

3. Explain what may be happening

Instead of immediately offering instructions, the AI first helps the parent understand the possible relationship between the data and the current difficulty.

4. Recommend a practical next step

The guidance focuses on a small, manageable adjustment rather than overwhelming the parent with a full programme.

5. Follow up and adapt

The product can ask whether the suggestion helped and use new information to refine future guidance.

How Numi’s AI learns

Numi combines retrieval-augmented generation with human-in-the-loop review to provide guidance that is personalised, grounded and continually improving. When a parent asks a question, Numi brings together relevant information from the child’s profile, recent sleep and feeding data, and a curated knowledge base informed by sleep research and specialist expertise. The AI uses this context to generate a response tailored to the child’s current pattern rather than relying on generic advice. Human review strengthens the system over time. Sleep specialists evaluate selected conversations, identify weak or unclear responses, and improve the underlying content, prompts, safety rules and retrieval logic. Parent feedback and follow-up outcomes also help reveal where the guidance is useful and where it needs refinement. This creates a continuous learning loop: Parent context → trusted knowledge retrieval → personalised AI guidance → human review → improved knowledge and system behaviour The goal is not to replace professional care, but to make everyday sleep guidance more relevant, understandable and responsible.

Leading from ambiguity to implementation

As both product lead and designer, I continuously balanced parent needs, product differentiation, AI capabilities and limitations, engineering complexity, budget, timelines, trust and long-term business potential — translating research and strategy into user flows, prioritised requirements and delivery trade-offs.

Working with engineers

I prepared flows, interface specifications and behavioural logic to clarify how the experience should function across different states. During implementation I reviewed builds, identified experience gaps and made trade-offs where the original concept was too complex for the initial release.

Working with specialists

I sought input from baby sleep specialists to improve the product’s guidance, language and boundaries — distinguishing between information that could be provided safely and situations that required professional escalation.

Roadmap leadership

I maintained a roadmap connecting feature development with the questions the product still needed to answer:

• Does personalised interpretation create more value than tracking alone?
• Which types of guidance lead to repeated use?
• How much information do parents need before they trust a recommendation?
• When is AI support sufficient, and when is human support more appropriate?
• Which features increase retention rather than only initial interest?
• What are parents willing to pay for?

AI value depends on product context

Conversational capability alone does not create a valuable AI experience. The quality depends on whether the AI has enough relevant context, communicates its reasoning clearly and helps the user make a meaningful decision. For Numi, value came from connecting conversation with recorded behaviour, developmental information and a clear follow-up loop.

Trust is a design outcome

Trust could not be added through a disclaimer alone. It was reflected in the language, recommendation structure, level of certainty, product boundaries and the relationship between data and guidance — resisting the temptation to make the AI sound more confident than the available information justified.

Focus is essential in zero-to-one work

The potential scope was much broader than what could be responsibly validated in the first release. Prioritising a focused guidance loop created a coherent MVP and made it possible to test the central proposition.

Leadership required moving between levels

I regularly moved between strategic, design and delivery decisions — defining the value proposition, then reviewing a microinteraction, clarifying logic with an engineer or deciding which feature to remove from the release.