Rhazes AI

Overview

Rhazes AI puts a clinician’s AI tooling in one place — an AI scribe, clinical decision support, automated medical coding, knowledge management, and auditing — instead of five disconnected products. I joined as a Full-Stack Engineer and now lead platform direction, with a focus on the thing that decides whether clinical AI is usable at all: whether a doctor can see where an answer came from and verify it.

That constraint drives the architecture. In healthcare, an unsourced answer isn’t a minor UX gap — it’s unusable. So most of my work sits on the seam between the model and the evidence behind it.

Key Features

• AI scribe and clinical documentation: Structured clinical notes generated from consultations, with reusable templates and drag-and-drop composition so documentation fits the workflow rather than fighting it.
• Grounded retrieval with traceable citations: I rebuilt the RAG layer around Page Index pipelines and reworked citation sourcing, so responses carry evidence a clinician can follow back to the passage that produced it.
• Audit workflows: Review and audit flows over generated output, giving teams a way to check the system’s work systematically instead of spot-checking.
• Automated medical coding: Coding support including post-coordination workflows, plus admin analytics for the operations side.
• Collaborative workspaces: Team invitations, role-based access control, and shared workflows so a practice works in one shared context.
• Subscription and entitlements: Freemium access tiers, feature gating, and scalable entitlement logic across the platform.

Technologies Used

• Next.js and TypeScript: Application and platform surface.
• LLM orchestration: Model workflows for scribing, decision support, and coding, with grounding and source attribution as first-class concerns.
• Retrieval-Augmented Generation: Page Index pipelines, chunk-level metadata, and citation-source mapping for traceable answers.
• Role-based access control and subscription architecture: Multi-tenant workspaces with entitlement-driven feature access.

Challenges and Learnings

The hardest problems were never about making a model produce fluent clinical text — they were about making its output reviewable. Rebuilding citation sourcing meant treating retrieval metadata as part of the product rather than an implementation detail of the vector store: if a chunk loses its section and page provenance during ingestion, no amount of prompt work recovers it downstream.

The UI/UX overhaul taught a parallel lesson. Redesigning the platform lifted user engagement roughly 20% and cut page load times about 30% — and much of that came from removing friction around reviewing AI output, not from adding capability. Trust in an AI system is largely a function of how cheap it is to check.

Outcome

Rhazes gives clinicians a single platform where AI assistance is auditable by default — every generated claim traceable, every workflow reviewable. I now lead architectural decisions across new capabilities, shaping how outputs are grounded, reviewed, and trusted as the platform grows.