Dataharvest 2026 - the European Investigative Journalism Conference

Most investigative workflows still rely on manually juggling dozens of tools. In this session, we'll walk through a live demo of a semi-automated pipeline built for real casework: web search and archiving with Playwright, face extraction, reverse image search, database cross-referencing with Telegram bots, social media analysis, and structured reporting via Obsidian mcp. All of this is orchestrated by Claude, an AI layer you can teach your own investigative methodology. At the end, participants will work through a simplified case using a workflow of their own.

Before the session, please install: Python, Claude Code. This session will teach participants to combine several smaller OSINT tools so they work together efficiently without requiring much manual effort. No special tools needed

"Agentic AI" is all the rage, but what does it offer beyond traditional LLM workflows? In this hands-on session we'll answer this question (and more) while leveraging Python's Pydantic AI library to build a start-to-finish agentic AI workflow.

Participants will learn how agents work, when they're useful, how to build custom tools, and options for tracing and evaluation. You'll leave able to write agentic workflows to extract information from texts, do semi-autonomous research, and deliver clean, structured results.

Basic experience with Python/LLMs is helpful but not required. After attending this session, participants will be able to understand when and how to apply agentic approaches to problems. Participants should have Python/Jupyter installed or a Google account for working in the cloud.

Projects involving FOI requests to multiple bodies often create significant challenges, from different file formats and data trapped in PDFs, to organisations providing data in different structures and different levels of detail. To get the big picture often requires data extraction, cleaning, reshaping, and checking.

In this session, we will share a series of tips and tools used to manage one project — including vibe coding with AI — which can be used to make any multi-response FOI project more efficient and accurate. No prior knowledge is required. By the end of this session, attendees should be able to design a data structure for an FOI project, use a range of tools, including AI, to extract, reshape, clean, and combine data from FOI responses, and design a data validation process to check AI outputs.

You will need a laptop with Google Drive and an account with an AI tool such as ChatGPT, Gemini, Claude, or Copilot. Installing Tabula and Open Refine will help you get more out of the session.

Tampermonkey is an age-old browser extension that allows you to inject scripts and stylesheets into any web page, turning the web into your personal playground. We'll look at how to customize your CMS with DIY features, add "Download all" buttons to paginated websites, automate tedious processes like filling out forms and redesign websites however you'd like. Best of all, Tampermonkey scripts are saveable and sharable, allowing you to give other members of your newsroom superpowers without fiddling with distributing extensions or asking them to run Python scripts. To follow along, participants should be able to install extensions in their web browser of choice.

Journalists often run into data that is visible on a website but impossible to download directly: a table buried in a government page, a list of public records, or search results that change with every query. Writing a full scraper can be time-consuming and technically demanding for what is often a one-time task.

This session introduces three lightweight approaches that cover most of these cases: reading a table directly from a page using pandas, downloading raw HTML and parsing it into a dataframe and pulling data through network requests. These techniques are practical tools for everyday newsroom situations. Participants will take home a GitHub repository with a working notebook to try on their own data, though some adaptation will be needed to apply it to different websites.

The three approaches vary in complexity. Basic Python knowledge is enough to follow along, but participants with more experience will be able to go further, and the code can be adapted with the help of an LLM.

You have a folder of documents and you want to extract data points from each one. And the data isn't in a structured table with neat rows and columns either. Here's where string functions and regular expressions can help. The demonstration will be in R but the skills are generic to all languages.

Say hello to Natural PDF, a new Python library for wrangling PDFs that's focused on usability and feature-completeness. Process PDFs with scraping-like selectors and spatially-aware queries, asking for "the red alphanumeric string" or "the content below the big Summary header." Beyond the basics, Natural PDF is also full of modern conveniences like table detection, multiple OCR engines, and citation-aware LLM data extraction.

To get the most out of this session, participants should have experience with Python and struggling with terrible PDFs.