Summary
- Engineered a complete content management system for clinical documents and their vector embeddings using Next.js and PostgreSQL (with pgvector extension and Cohere API).
- Successfully merged the standalone admin project into the primary Clinical Audit Platform to serve as its core content management panel.
- Improved core platform functionality by implementing dynamic data fetching with pagination and ensuring AI prompts included relevant clinical guideline search results.
- Refined the user interface by fixing display issues, updating dashboard components, and correcting layout inconsistencies.
- Contributed to the project’s stability by fixing data structure bugs, resolving linting warnings, and improving application performance.
- Worked on a platform utilizing an enhanced Retrieval-Augmented Generation (RAG) system for multi-source medical literature searches with domain-aware filtering.
I joined the Clinical Audit Platform project midway through its development, contributing to a sophisticated SaaS platform designed for specialist clinicians. The platform provides a structured, 6-step workflow for clinical audits, featuring an advanced AI-powered literature search and multi-tenant pharmaceutical company sponsorship.
My work was divided into two main phases: first, developing a standalone content management system, and second, integrating it into the main platform and subsequently enhancing existing functionalities.
Phase 1: The Clinical Embed Admin System
My initial task was to build a full-stack administrative dashboard to manage the clinical data embeddings that power the platform’s AI search.
This standalone project, the Clinical Embed Admin, was architected using a modern tech stack centered around Next.js and TypeScript. The backend utilized a PostgreSQL database with the pgvector extension for efficient similarity searches, managed by the Prisma ORM.
A guideline can have a source document and multiple chunks, and each chunks can have their own facts and concepts, which summerized and distilled from one part of the source document. When generating the embedding, the system first extracts the core, searchable text from a chunk, including its content and associated facts. This consolidated text is then sent to the Cohere API to produce a vector representation. This vector, along with rich metadata about the chunk and its parent guideline, is stored in a specialized database table (with a column of vector datatype, from pgvector extension). Base on chunk last modified time and embed generating time, each chunk will display a status from [✓ Embedded, ! Outdated, ✗ No embedding] in front end.
This system handled the complete workflow for clinical documents: uploading, storing original chunks, generating vector embeddings, and storing them for retrieval. It provideded the essential foundation for the platform’s content management capabilities.
Phase 2: Integration and Platform Enhancement
Following the successful development of the admin system, I merged it into the main Clinical Audit Platform to serve as its integrated content management panel.
Once the integration was complete, my focus shifted to testing, fixing, and improving the platform’s existing features. This enhancement phase involved a variety of tasks, from significant functional upgrades to minor UI fixes. Key improvements included reimplementing the data filter function with dynamic fetching and adding pagination to handle large datasets effectively.
I also worked on refining the AI system by ensuring guideline search results were correctly included in the prompts for clinician reflection.
Additionally, I addressed various bugs, fixed UI display issues, updated data schemas, and helped eliminate linting warnings to improve overall code quality.