Back to personal projectsHarman Singh
AI-led personal project

RxAssist: taking a doctor’s workflow from need to working software

A personal product build covering prescription creation, printing and an AI-assisted review.

AI used
Claude CodeLLM API
Built with
ReactFastAPISQLite

Project brief

A doctor in solo practice needed a prescription workflow that fitted the consultation and printed on a familiar pad layout. I scoped and built RxAssist with Claude Code: an application that runs on the doctor’s laptop for entering consultation details, creating a prescription and generating its PDF. I also added an AI cross-check as a separate capability for the doctor to assess, while keeping the core prescription workflow usable without AI.

My role

I handled product scoping, workflow and design decisions, AI-assisted development, testing and preparation for the doctor’s review. The doctor provides the clinical judgement.

467
automated tests passing

Key decisions

  1. 01

    Keep prescription creation independent of AI

    A model error or an unavailable service should not stop the doctor from creating a prescription. I separated the AI review from the form and PDF workflow, so the core task works by itself.

  2. 02

    Stop the handover when the audit found defects

    An independent audit found that long prescriptions could lose medicines in the PDF and that AI requests were failing. I paused the handover and organised the fixes into gated releases, checking the output instead of trusting that the build looked complete.

  3. 03

    Keep the AI review within a clear boundary

    Patient records are stored on the laptop. The AI request carries selected clinical details without the patient’s name, and its response appears separately for the doctor to assess. The model does not make the prescribing decision, and its response is not printed on the prescription.

Outcome

Built the application, with 467 automated tests passing, and prepared the package for the doctor’s review. The doctor’s review of the clinical reference list, the AI cross-check and the pad layout comes before regular use.

AI and tools

Claude Code supported development. An LLM API powers the optional prescription cross-check. I set the workflow, reviewed the build and decided what needed further testing before use.

Get in touch.

Building something in FinTech, banking or consumer products? I’d like to hear about it.

Email meLinkedIn →Resume (PDF)