Writing depends on finding the right knowledge
Clinical-trial writing brings together research material, specialist terminology, previous documents, and detailed requirements. The challenge is not simply producing more text. Writers need to locate relevant information, understand its context, and turn it into a document that remains subject to expert review.
For TrialCraft, Gilles worked as technical lead for product and AI on an LLM-based clinical writing and compliance assistant. The remit included technical roadmapping, funding pitches and proposals, cloud infrastructure, and implementation of the frontend and backend. The engagement joined product direction with hands-on application engineering.
A retrieval-led conversational foundation
The assistant combined conversational interaction with retrieval-augmented generation. Its backend connected document ingestion and retrieval with a service interface for the application. That foundation supports a focused workflow: bring relevant material into context before asking a model to help with a writing task.
For specialist document work, retrieval quality and generation quality answer different questions. Did the system find material relevant to the user's request? Did the response represent that material appropriately? Treating those as separate engineering concerns gives the team a more useful way to reason about a disappointing answer.
Product decisions and engineering decisions together
Technical leadership meant connecting an ambitious product idea with an implementable application. The engagement covered both the user-facing experience and the services underneath it, alongside the cloud foundation needed to run them. Roadmapping and funding work required the same translation from specialist needs into concrete capabilities and delivery priorities.
This kind of product needs deliberate boundaries. A writing assistant can support information access and drafting, while clinical and regulatory judgement remains a responsibility for qualified professionals. Clear task definitions help distinguish useful assistance from claims the application cannot establish, including any suggestion that generated text itself proves compliance.
A coherent foundation for specialist assistance
The contribution brought product direction, RAG engineering, frontend and backend implementation, and infrastructure into the same technical remit. Working across those layers connected the writing task to the software needed to support it, from the information available to the assistant through to the experience of asking it a question.
The transferable value is a way of approaching high-context AI products. Start with the expert's task, make the knowledge layer explicit, and connect the assistant to an application the expert can actually use. That gives development a clearer purpose than a general chatbot and a more practical basis for reviewing the system's limitations.
