Yousra Diab AI Engineer

Years of political agendas sat in a meeting system with no useful search. I built the Go integration that collects every agenda as it is released, and a Copilot Studio agent that answers questions about agendas and meetings through Azure AI Search.

The client would not put case data in the cloud, so I built the two assistants on an open model running on its own servers, with Qdrant for search and Tika for documents: one finds earlier decisions in years of agendas and minutes and answers with a timeline and its sources, the other reviews a draft case presentation inline, on structure, wording or language, without rewriting a word.

Every decision letter was typed into the case system field by field. I built a Go service that reads each letter, fills the fields and flags what it is unsure of: two model calls per letter, with the medical terms matched against three reference vocabularies in a vector index. The flagged fields go through the review step the client required for compliance.

An ageing R script was run by hand and its results copied into an Excel workbook and a presentation. I rebuilt the pipeline in Go as a scheduled Docker job: it applies the four substitution rules, fills the workbook in place without touching the hand-maintained columns, and generates the presentation.

The same questions were answered over and over from a knowledge base nobody could search well. I built an Azure Function that syncs the articles into Azure AI Search, and a Foundry agent, triggered from the case system, that reads a new case, finds the relevant articles and proposes the next step.

The client wanted to ask questions across its internal SharePoint drives, but not everyone may read everything there. I built the retrieval on Azure AI Search with document-level access control: the SharePoint indexer ingests the existing file and folder permissions with each document, and every query carries the identity of the person asking, so the index itself trims the results. A document nobody has shared with them stays invisible, even to the model.

Municipalities wanted to build their own agents, not only have them built. I wrote the teaching material for Copilot Studio and ran it in two formats: a one-day introduction that ends with an agent built together, and a course that meets every third week and builds on the last session.

Every call to a public phone line went through to a person, including questions the knowledge base could answer. I built the routing for the NORA voicebot: it works out what the caller is asking, answers from the knowledge base when it can, and transfers by intent when it cannot, over a blind SIP transfer. If the transfer fails, the caller is still on the line.