AI Research Insights Platform

The Problem: Making a RAG system work in a demo is relatively straightforward. Making it work reliably in production — with real users, real queries, variable data quality, and the need for accurate, source-grounded answers — is significantly harder.

The Solution: The retrieval layer uses GCP Discovery Engine, Google Cloud's managed vector search and document understanding service. Survey responses are indexed as structured documents, enabling semantic search at scale. Primotly maintains the indexing pipeline and query layer. The LLM synthesis step uses a large language model to generate the summary from retrieved context. Prompt engineering ensures that summaries are grounded in the actual retrieved data rather than generating information not supported by the surveys.

The Result: users receive answers that are traceable to specific survey responses, with brand and keyword attribution visible in the result set.

The Problem: The platform was built on a legacy in-house platform that had accumulated significant constraints. The entire application — not just the frontend — needed to be moved to a new stack without disrupting the existing AI data pipeline or active users.

The Solution: Primotly is executing a full rebuild rather than an incremental migration. The frontend is being built from scratch in Angular. The backend is being rebuilt in Python with clean separation of concerns. The GCP infrastructure — Discovery Engine, Firestore, and the LLM integration layer — is being re-wired to the new application. What can be preserved from the previous implementation is carried forward; what cannot is rebuilt to a higher standard.

The Result: a fully owned, modern application on GCP — free from legacy platform dependencies, with a maintainable codebase and a clear path for future development. Migration is currently in progress.

The Problem: The end users are not data scientists — they are brand managers and marketing professionals who want answers, not data. The interface needed to feel as natural and fast as a search engine while surfacing the depth of information that survey data contains.

The Solution: The new frontend is designed around the brand manager's actual workflow: type a question, get a summary, explore supporting evidence. The tagging system allows users to quickly filter results by brand, topic, or attribute without reformulating their query. The interface prioritizes readability and scanability over data density.

The Result: a customer-facing interface that lowers the barrier to accessing research insights — making the survey data as accessible as a search engine, not as opaque as a report archive.