← All selected work

Ishara: Interpreter Dispatch Agent

Books a sign language interpreter from a forwarded appointment message, in English or Arabic

Live demo with fictional interpreters · 2026 Source Live demo
ishara-tan.vercel.appOpen full size ↗Ishara review page showing an Arabic hospital appointment message parsed into location, area, date, visit type and sign language fields, with a Confirm request button

The problem

Deaf residents in the UAE often have to arrange a sign language interpreter themselves for hospital, government and school appointments. The appointment details are usually in a message they already received. The project asks whether an agent can turn that message into a booked interpreter, while keeping the resident in control.

My contribution

Personal project, built for the Dubai Chamber Create AI Agents Championship (societal category): message parsing in English and Arabic, the matching and replacement agent, the requester and coordinator interfaces, and deployment.

  • FastAPI
  • SQLite
  • Next.js
  • Gemini (OpenAI-compatible)
  • Railway
  • Vercel

What I built

The approach

A Deaf resident pastes an appointment SMS or email. The agent reads it, asks the resident to confirm the details, offers the slot to qualified interpreters in waves, keeps a backup ready, and finds a replacement automatically if the interpreter cancels.

What the system guarantees

0interface languages, with right-to-left Arabic
0offers sent before the resident approves
0API regression tests

The live demo uses fictional interpreters and simulated replies. It has not yet been tested with Deaf residents, interpreters or Braille displays; those trials are the next step.

Follow the workflow

The resident pastes an SMS, email or WhatsApp message about an appointment, in English or Arabic.

More from the project

Select an image to inspect it at full size.

Engineering decisions

01

Rules first, model second

A rules parser reads English and Arabic messages on its own. A hosted model only refines it, and any model failure falls back to the rules reading.

02

The resident approves before anyone is contacted

Parsed details are shown for correction, and fields the agent guessed are marked. No interpreter sees the request until the resident confirms.

03

Plan for cancellations

Each outcome updates an interpreter reliability score used by the next match, and a cancellation triggers an automatic replacement.

Limitations & next steps

Run a pilot with Deaf residents and certified interpreters, record a reviewed Emirati Sign Language confirmation video, and test the requester page with Braille displays.