Idiap Research Institute

William DROZ

Research & Development Engineer

Discover CPAI GitHub Publications
Portrait of William Droz

Also at Idiap

BISS

Bugzilla and Intranet Semantic Search

I develop BISS to help colleagues find information across Idiap's Bugzilla, technical documentation, and BookStack. It combines semantic search with AI-assisted answers that link back to their sources, through the website, a terminal client, and Mattermost bots.

Open BISS Internal Idiap service
How BISS works Subprojects, shared search & Mattermost assistants

From internal knowledge to useful answers

  1. 01 · Collect

    Dedicated scrapers gather discussions, documentation, and BookStack content, preserving the original text, dates, and source links.

  2. 02 · Retrieve

    BISS indexes the content as passages and combines matches by meaning and by keyword to find relevant evidence for a question.

  3. 03 · Answer

    Users can explore the search results or ask an AI assistant to build an answer from the retrieved evidence, with links to its sources.

The subprojects and their roles

biss-data-scraperBugzilla collection
A Python scraper that collects issues and their discussions, supports incremental updates, and sends them to BISS in a common document format.
web_scraperTechnical documentation
A Rust crawler that extracts documentation pages, titles, and source links, then sends them to the same ingestion service.
bookstack_scraperBookStack pages & attachments
A Python collector that reads BookStack pages through its API and extracts text from supported attachments, including PDFs and Office documents.
biss_webThe shared knowledge service
The Rust application: Dioxus for the interface, Axum for the API, Qdrant for retrieval, MongoDB for document storage, and Rig for answer generation.
BISS MCP · part of biss_webTools for AI assistants
Exposes search and source-reading tools through the Model Context Protocol (MCP), connecting assistants to the same BISS knowledge service.
biss-cliBISS in the terminal
A Rust client with command-line and interactive modes to search, filter results, inspect source text, and ask questions without leaving the terminal.
mattermost-botsAnswers in chat
Routes questions to specialist assistants, including BISS for focused lookups and DeepBISS for investigations that need several searches and evidence review.
fakellmServing the agents
Exposes the agents through an OpenAI-compatible API so Mattermost and other compatible clients can use the same workflows.

How the Mattermost assistants use BISS

  1. Mattermost
  2. mattermost-bots / fakellm
  3. BISS or DeepBISS
  4. MCP tools
  5. BISS knowledge service

The BISS assistant searches and reads the original source text to answer a focused question. DeepBISS first plans an investigation, runs a primary search and a second research pass to cross-check the evidence, then asks a reviewer agent to produce the answer with source links.

The wider Mattermost bot collection also includes specialists for CPAI, Idiap publications, room availability and booking, and Slurm computing jobs, each connected to its own relevant tools or data.

Short biography

I've been programming since 2004 and received my master's degree in software engineering in 2015. I worked as a research assistant at He-arc Ingénierie until 2018.

I'm now working as a Research and Development Engineer at Idiap, where I work with technologies including PyTorch, TensorFlow, and Spark.

My latest major project is CPAI, a web platform supporting public participation in AI research at Idiap.

Skills

  • Machine Learning
  • Python
  • Databases (SQL & NoSQL)
  • git / bash / Linux / Docker / ssh / gRPC
  • Distributed Computing
  • Optimization
  • Back-end

Sharing practical AI

Teaching & workshops

I teach students, workshop participants, and colleagues how to build with language models and AI agents, through practical exercises and real applications.

Teaching · HEVS

Applied LLMs for Life Sciences

Life science students · HES-SO Valais-Wallis

A 14-week course covering AI-assisted Python development and reliable LLM applications. Students build agents with structured outputs, tools, and retrieval, with an emphasis on testing, reproducibility, and human review.

Course materials
Workshop · AI Days 2026

Getting Started with AI Agents in Python

· Martigny · In French

A hands-on introduction to building AI agents in Python, with exercises in data analysis and visualization, tool calling, connecting tools through the Model Context Protocol (MCP), and browser automation.

Workshop materials Event GitHub
Workshop · ICC 2026

LLMs At Your Service: How to Use Them Efficiently?

Idiap Create Challenge · · Martigny

A workshop for Idiap Create Challenge participants on using LLMs to build prototypes. Topics include AI coding tools, agent workflows, writing specifications, and verifying the results.

Workshop materials Event
Internal workshop · Idiap

AI Agents from Scratch for Idiapers

Idiap colleagues · Around 30 attendees

A hands-on introduction to building Python AI agents from scratch. Exercises covered structured outputs, data analysis and visualization, controlled tools, and connecting agents through the Model Context Protocol (MCP).

Open source

Ongoing projects

Personal project

noabspath

Detect hard-coded absolute paths in your codebases.

View on GitHub Personal project

check_urls

Verify the validity of URLs inside your files.

View on GitHub Idiap project

FakeLLM

Serve PydanticAI agents through an OpenAI-compatible API, so existing clients can interact with agent workflows.

View on GitHub

Research & development

Work projects

WeNet

Develop the culture, science, and engineering methodologies, algorithms, and social interaction protocols for an online platform that will empower machine-mediated diversity-aware interactions between people.

Scientific work

Publications

Get in touch

Contact

Email
william.droz@idiap.ch
Telephone
+41 27 721 77 62
Office
408