From failed vendor to live product in 30 days

The company

We CU is a patient advocacy organization for people living with chronic urticaria, commonly known as chronic hives. The organization operates under the Global Allergy and Airways Patient Platform (GAAPP) with sponsorship from Novartis, one of the largest pharmaceutical companies in the world.

Chronic urticaria affects millions of people globally, but it doesn't get the attention or funding of higher-profile conditions. Patients often struggle to find reliable, plain-language information about their disease, its triggers, treatment options, and what to expect long-term. We CU exists to close that gap.

The problem

We CU and Novartis wanted an AI-powered chatbot called Scratch: 24/7 patient access to medically vetted information about chronic urticaria. The concept was straightforward. Take decades of published medical research, load it into a system that could answer patient questions conversationally, and make it freely available online.

They hired a firm with domain expertise in the disease area to build it. Over 12 months later, the project still wasn't finished. When the team finally pushed a beta live, the application went down within days and stayed down for over 30 days once real user traffic started flowing.

The approach

We were already a partner to the GAAPP team through other projects. When the Scratch build stalled, GAAPP brought us in to take over.

We didn't start from scratch on the research. The medical content had already been assembled and certified: 174 articles written and reviewed by professionals in the chronic urticaria space. The raw material was solid. The problem was that no one had been able to turn it into a working product.

We assessed the existing state of the project, identified the core technical failures, and started building. No multi-month discovery phase. No steering committee. Get into the problem fast, build against real data, ship something that works.

The solution

The system has four layers:

1. The knowledge base. All 174 medically certified articles vectorized into an indexed knowledge base. Scratch uses retrieval-augmented generation to pull relevant passages and construct accurate, source-backed responses. Every answer traces to content approved by the medical advisory board.
2. The recency engine. Some older articles were directly contradicted by newer publications. A prioritization layer automatically weights newer research over older when conflicts arise, like a newsroom editor going with the fresher source.
3. The agent network. Not a single chatbot but a team of specialized agents: a knowledge agent for core Q&A, a policy agent keeping conversations on-topic and declining off-scope questions, a translator agent supporting 15+ languages, and a vision agent for uploaded documents and images.
4. The infrastructure. Architected for production traffic from day one. No 30-day outages, no beta-stage crashes. Real users, at scale, reliably.

What came next

Scratch is now live and serving hundreds of thousands of patients globally. Novartis continues to back the project as part of their commitment to patient education in underserved disease areas. People living with chronic urticaria get a free, always-available resource for understanding their condition, without waiting for a doctor's appointment or sifting through unreliable search results.

Since then, we've built Pulmo, a global asthma chatbot using the same architecture, and the Asthma Canada bot. GAAPP has engaged us to build eight additional disease-specific chatbots across their other focus areas, and we're actively building chatbot solutions with another major pharmaceutical sponsor. What started as a rescue mission for a failed vendor project became the foundation for a multiyear partnership across one of the largest patient advocacy platforms in the world.