Skip to content
All projects

Client work / Hakim Click

Hakim Click

A clinical assistant for doctors in Ethiopia that answers questions only from 38 national treatment guidelines and links each statement to the exact page it came from. It also transcribes consultations and drafts notes the doctor signs off.

38 guidelines, 8,076 pages and 13,536 indexed passages, and every answer links back to the page.

Clinical AI for notes, dictation and cited answers in five languages
Every sentence carries a citation to the guideline page it came from
A copilot raises red flags from the live transcript
The consultation becomes a SOAP note that stays a draft until signed
Sign-in with seeded demo accounts for each plan
Installable on a phone

01 / 07Clinical AI for notes, dictation and cited answers in five languages

Overview

Doctors in Ethiopia are expected to follow the national treatment guidelines, but those guidelines are spread across dozens of long PDFs. Finding the right dose in the middle of a clinic day means scrolling through hundreds of pages, and a general chatbot might invent an answer that sounds right.

Hakim Click only answers from the guidelines themselves: the Standard Treatment Guidelines, the malaria, neonatal, obstetric and NCD protocols and others, 38 documents in all. Each sentence in an answer carries a small citation, and clicking it opens the cited page with the passage highlighted, so the doctor can check the claim without leaving the conversation.

The product has two parts. A Next.js app handles accounts, plans, consultations and chat. A separate Python service does the retrieval, answer generation and grounding checks. Binyam built both.

Features

  • Ask a clinical question and get a short answer where every statement cites a guideline page
  • Open a citation to see the source page with the passage highlighted, or open the full PDF at that page
  • Record a consultation in Amharic, Afaan Oromo, Tigrinya, Somali or English and read the transcript with an English line beside it
  • Turn a consultation or dictation into a SOAP note, summary or referral letter that stays a draft until the doctor signs it
  • Pick a plan with a daily credit allowance and see what the day's credits were spent on
  • Install it on a phone as a web app that keeps working through a dropped connection

Challenges

  • A wrong dose is a patient-safety problem, not a bad search result. Retrieval combines vector and full-text search, a language model reranks the top candidates for direct relevance, and the assistant declines to answer when nothing in the guidelines supports one.
  • Dosing charts and triage grids lose their meaning when split into plain text. Ingestion keeps 2,352 tables as markdown with their column headers intact, so a dose stays attached to its weight band.
  • Answers take 7 to 25 seconds, which runs into serverless time limits if proxied. The browser streams straight from the retrieval service using a 15-minute token that carries only a user id and plan, with no personal data.
  • Linking to a PDF page can scroll to it but cannot show which paragraph was used. The service renders the cited page as an image and marks the passage the answer drew from.