Week 40: Digit Conference 2026

This year’s Digit Conference is hold at kammivabrik in Tartu. From my understanding, it is an event majorly for python developers. In the morning at 5 a.m. I got up and took off to catch the bus at Tallinn Bus station, and I left Tartu Bus station at 8 p.m.

The conference includes two sections, software engineering and data & machine-learning. They all includes workshops to attend.

Before I go, I made a plan to attend these topics.

• Synthetic Counterfactuals for Text Classification: Detecting AI-Written Job Postings in the Wild
• What We Can Learn from Exemplary Python Documentation
• MLVC: Building the World’s Most Efficient Video Codec
• Building Kubernetes-Native Python Applications
• Own Your LLM: Lessons from Building a Local AI Agent
• You Are the Harness. Until You Build One
• Causal Agentic Data Science

The speeches are in the same hall, so one of it has to wear headphones.

What We Can Learn from Exemplary Python Documentation

This is a workshop by Christian Heitzmann, aims to use Sphinx to generate code docs into html. The document structure uses reStructuredText as markup text structure.

The workshop also introduced google style docstring. kroki is used as diagram tool, however, I went to another speech in half of the workshop.

MLVC: Building the World’s Most Efficient Video Codec

This is a speech by Evgenii Indenbom. This codec is used for microsoft teams, which is useing numberical regression to make predict frames to insert inside of actual frames. The paper is also a worth to read.

Building Kubernetes-Native Python Applications

This is a workshop by Lauri Võsandi. The project is to use MCP to develop new features on lolcatz. Well, the original repo needs to setup kafka credential, can use mcp to make it done.

Own Your LLM: Lessons from Building a Local AI Agent

This is a speech by Heili Aavola, who was using local LLM agent (Hermes agent) with 2 Nvidia GPUs.

Although, I do not think local LLMs can solve all the security issues(should be harness). I really like the cheatsheet. It compared several runtimes, vLLM & tensorfold, several local runable models and some easy configs for spark platform.

You Are the Harness. Until You Build One

This is a speech by Kirill Linnik. The topic is about multi-agent. He introduced OpenAI Symphony which can be used for agents orchestration. Simply speaking, the agents will do the issue fixing and validation work instead of human.

He also mentioned DevClub, a monthly meetup event in Estonia.

Causal Agentic Data Science

This is the last one, the speech is by Carlos Trujillo. I thought it would about some data science. But it was more like the workflow. He mentioned paseo. I have never used it befor. I intend to use it.