Skip to content
Meftun Akarsu
LLM systems & agents · 2026

EAA & GDS chatbots

Copilot Studio · Azure AI Search · WCAG · Evaluation

Two live Copilot Studio chatbots for an international accessibility organization, with accessible chat pages, a search pipeline and an evaluation set.

Rows of raised braille dots on a light gray plate, seen at a steep angle against a dark background
Photo: Ramona / Unsplash

Overview

Two chatbots on Microsoft Copilot Studio that I built and maintain at RADIAITE for an international accessibility organization. Both are live: one answers from the organization’s knowledge base, the other searches a public database of commitments. People reach them through accessible chat pages embedded in the organization’s website.

What I built

I wrote the chat pages: inline chats whose start screen carries a privacy notice, and a floating widget. For the commitments bot I built the Python pipeline that indexes the database in Azure AI Search, and I maintain the agent’s instructions, so that it answers in the language of each message rather than the browser’s. Power Automate flows take new knowledge-base entries in and republish the agent.

Accessibility and evaluation

Two WCAG audits found no axe violations before or after a chat starts and no horizontal overflow at 320 pixels, and the link color reaches a contrast of 5.88:1. A set of 52 questions compares the classic agent with its successor before a change goes live. One lesson stuck: a widget that passed three automated audits froze the whole page once the website’s CMS nested it differently, so every embed is now tested inside a real page.

Private codebase

Built with

Copilot Studio · Power Automate · Dataverse · Azure AI Search · Python · HTML · axe-core

← All projects