Led strategy and design for a mobile app and AI companion built to support people with T1D as they navigate the complex journey from diagnosis through treatment to cure. Clients: Fueled/BT1D
We were engaged to create the product strategy and technical design for a first of its kind mobile app leveraging AI alongside the supportive services and content already native to the web experience of this global health charity.
I led a strategy and design engagement defining a net-new mobile product centring an AI-native experience in a community-saturated competitive landscape.
There are 9.5 million+ people living with type 1 diabetes worldwide, and that number is climbing — up roughly 13% since 2021. For most people a T1D diagnosis arrives out of nowhere and carries an immense emotional burden and significant cognitive weight.
No two individuals share the same journey. Someone three days post-diagnosis and someone ten years into treatment are very different users who may access the same platform, but require a unique and personalized experience. They need different things, in different tones, at different moments. Most products in this space fail to deliver the level of personalization and support afforded by modern AI architectures. The market is community-saturated and personalization-empty, competing primarily on features such as forums, feeds, and content. Almost nobody leads with an experience that adapts to the specific human using it.
TLDR:
We bet on a personalized AI companion rather than competing on community. The vision for the companion was simple: be present across the entire experience in a way that is native, subtle, and personal. It is truly your personal and adaptive guide woven into the entire experience, but also accessible through a dedicated space where you can talk, journal, and reflect. It runs onboarding, curates what you see day to day, and points you toward the right content and next step in treatment.
I anchored personalization on three core axes: who you are (your unique persona), why you're here (as a patient, advocate, fundraiser), and time since diagnosis (for you or the loved one in your life). People at different stages require a different guide, with each suggestion grounded in the user's unique position in their journey with the disease. The organization has a rich history and a robust set of services already — events, community, and vetted expert content. The app routes to what's already delivered by the organization through a set of smart digital doorways leading to the appropriate resources. The aim was for V1 to build a relationship of trust by being useful and personal first, with the heavy clinical offerings coming in later build phases.
TLDR:
I designed a multi-model architecture to provide personalization to the user and adaptation to the user's cognitive and emotional states when interacting with the app, along with a phased approach to the agent's memory. Claude Sonnet is the primary, Claude Haiku handles classification and emotional-state grading, and OpenAI GPT sits as a warm fallback for availability. Sonnet follows nuanced instructions, refuses medical advice cleanly while staying warm, and modulates tone; Haiku is fast and cheap enough to run on every turn; GPT covers an Anthropic outage.
The LLM is never called directly from the mobile client. Keys, rate limits, prompt-injection defense, cost, observability, PII scrubbing, moderation, and evaluation all live on the backend. The app talks to an orchestrator; the orchestrator talks to the models.
WordPress stays the content source of truth. A scheduled REST pull normalizes content to Markdown, chunks it (~500 tokens, 50 overlap, section-aware), and embeds it into pgvector. Each chunk carries metadata: source URL, last-modified, content type, provenance (expert-verified / org-vetted / community-contributed), and locale. Retrieval is hybrid — pgvector ANN plus Postgres tsvector full-text, reranked, top-8 into the prompt. Library-mode answers cite the WordPress sources beneath the response; if retrieved content doesn't cover the question, the AI says so rather than inventing an answer.
Session-only at MVP (15-minute idle timeout, no raw transcripts in the database). At v1, opt-in persistent memory as a compressed summary rebuilt every N turns and capped ~500 tokens, which the user can view and correct in the Account tab; raw transcripts kept 30 days max, only if opted in. Context handled by rolling summarization at ~70% of the model's context limit.
A ~100-prompt golden set curated with the organization's content and clinical advisory covers content Q&A, wayfinding, crisis signals, out-of-scope, pediatric, prompt injection, and T1D/T2D disambiguation. Metrics: retrieval precision@5, answer faithfulness (LLM-as-judge), safety pass/fail, tone match. Regression runs on every prompt or model change, with deploys blocked on safety regressions. Modeled cost is ~$1,400/month at 10K MAU on a 70/25/5 light/medium/heavy split, held down by per-user daily caps and pushing classification to Haiku; the GPT fallback at ~10% is roughly cost-neutral.
The scope of this engagement covered strategy and design, and the architecture I proposed is set to deliver a truly unique and personalized mobile experience for this important global community.
Engagement led at Fueled. Some product and client details are intentionally held while the work is unreleased.