PollenPilot

AI-powered Hay Fever Assistant

Problem

Melbourne's extreme seasonal pollen conditions create significant challenges for allergy sufferers, with limited tools available for personalised, contextual management. Existing solutions provide generic advice without considering individual circumstances, activities, or real-time environmental factors.

Key Hypotheses

• Hay fever sufferers prefer personalised recommendations delivered in natural language, specific to their context and activities

• Proactive alerts are essential on high-risk days when environmental conditions are deteriorating

Research

User Research Foundation

I conducted interviews with chronic hayfever sufferers to understand current management practices and identify key gaps:

• Trust through transparency: Users lost confidence when they couldn't correlate their symptoms with app recommendations

• Personalisation failures: Existing apps provide generic advice despite individual sensitivities to different allergens

• Activity-specific guidance: Experienced sufferers understand when conditions are "bad" but lack granular weather and pollen correlation data to optimize timing for outdoor activities

Prototype Development

Vision & Technical Strategy

Original Vision: Native mobile app with real-time API integration to weather and pollen data, featuring an AI assistant accessible via text or voice interface.

Strategic Pivots for Rapid Testing:

• Timing reality check: Since it wasn't pollen season, real-time data lacked meaningful variation—pivoted to mock data scenarios for different environmental conditions

• Platform decision: Built web app instead of native mobile for faster testing cycles

• Interaction focus: Prioritised text/chat interface over voice to validate core hypothesis first

Technical Implementation

Core Architecture:

• Platform: Web application using Claude API for rapid prototyping and iteration

• Data strategy: Mock environmental scenarios based on historical Melbourne patterns

• AI Training: Developed core interaction patterns as training examples:

  • Morning planning consultations

  • Real-time activity decision making

  • Multi-day planning ahead

  • Reactive symptom management

Melbourne-Specific Knowledge Integration: Built comprehensive local context into system prompts:

• Seasonal patterns: Local pollen seasons and peak timing windows

• Botanical specificity: Melbourne's problematic grass types and their bloom cycles

• Wind dynamics: Local wind patterns and their effects on pollen distribution

• Weather phenomena: Thunderstorm asthma patterns unique to Melbourne's geography

Key Insights

I recruited 5 participants with varying severity of allergies to validate the core hypothesis for the product. The testing focused on core user flow and AI interaction patterns.

User Feedback Summary

• 2 out of 5 users represented the core target demographic

• Educational value emerged as an unexpected benefit

• Personalisation resonated strongly with users

• Voice interaction showed promise for hands-free use cases

Critical Learnings

• AI-powered products aren't universal: some users prefer traditional UI interactions

• Natural language recommendations create stronger engagement than generic alerts

• Contextual awareness (calendar integration) significantly improves recommendation relevance

Lessons Learned

AI Product Development

• Reliability vs. Innovation: Balance cutting-edge AI capabilities with consistent user experience

• AI as collaborator: Using AI as a thinking partner accelerated both design and development processes

• User reference diversity: Not all users embrace AI interaction, suggesting the need to provide multiple interaction modes

Process OptimiSation

• Strategic Descoping: Focused on rule-based recommendations over complex predictive models

• Parallel Workflows: Maintained simultaneous building, recruiting, and testing threads

• Flexible Timeline: Adapted sprint structure based on learning velocity

Impact & Future Opportunities

This project demonstrated the potential for AI-powered, contextual health applications while revealing important constraints around user adoption and technical implementation. The rapid development cycle proved that meaningful AI products can be built quickly with the right tools and approach.