DataTalksClub

Free online community for data builders.

Learn data, ML, MLOps, LLMs, RAG, and AI engineering.

Join free hands-on courses, events, workshops, and podcasts.

[DataTalksClub @DataTalksClub]

Want to talk about data?

Data science, career questions, practical machine learning, and more?

Join @DataTalksClub! The link is in the profile description.

Learn more about our activities in the thread! 🧵

[DataTalksClub @DataTalksClub]

AI agents are inflating the numbers you rely on.

🔸 Engagement metrics inflated by machine traffic
🔸 Ad attribution quietly broken
🔸 One industry hit harder than most

Here's what to do about it 👉 https://t.co/PTMIpDYw3O

[DataTalksClub @DataTalksClub]

Dewang Sultania suggests a guardrail check: list every tool your agent can call, then ask what happens if each runs incorrectly 100 times. Start with Chapter 1 of Designing AI Systems.

https://t.co/Ay5TB0C6Nd

[DataTalksClub @DataTalksClub]

AI agents are inflating the numbers you rely on.

🔸 Engagement metrics inflated by machine traffic
🔸 Ad attribution quietly broken
🔸 One industry hit harder than most

Here's what to do about it 👉 https://t.co/PTMIpDYw3O

[DataTalksClub @DataTalksClub]

Dewang Sultania suggests a guardrail check: list every tool your agent can call, then ask what happens if each runs incorrectly 100 times. Start with Chapter 1 of Designing AI Systems.

https://t.co/Ay5TB0C6Nd

[DataTalksClub @DataTalksClub]

What helps AI agents answer business questions correctly?

Next Query 2026 by PhoenixAI | Nov 4 | Free virtual summit

➡ Business meaning + fresh context
➡ Permissions + resilient analytics
➡ Production architecture + migration lessons

Register: https://t.co/MLNLz3tI30

[DataTalksClub @DataTalksClub]

AI agents are inflating the numbers you rely on.

🔸 Engagement metrics inflated by machine traffic
🔸 Ad attribution quietly broken
🔸 One industry hit harder than most

Here's what to do about it
👉 https://t.co/PTMIpDYw3O

[DataTalksClub @DataTalksClub]

https://t.co/kFn2eOwYY5 Weekly #299 is out.

This week: AI agent traffic vs real readers, ML Zoomcamp Module 5, Project 1 criteria, two Nov 2 events, and a Tejas Kumar recording on AI evals.

https://t.co/swcV9sCPRL

[DataTalksClub @DataTalksClub]

https://t.co/kFn2eOwYY5 Weekly #299 is out.

This week: AI agent traffic vs real readers, ML Zoomcamp Module 5, Project 1 criteria, two Nov 2 events, and a Tejas Kumar recording on AI evals.

https://t.co/swcV9sCPRL

[DataTalksClub @DataTalksClub]

Don't miss our next live podcast episode! We're excited to have Louis-François Bouchard join us to discuss Shipping Reliable LLMs and the Rise of AI Engineering. Tune in to learn more about their experience and insights.

Register here: https://t.co/PvZUIlIh6J

[DataTalksClub @DataTalksClub]

In our webinar GraphRAG in Real-Life Use Cases, Daan will:

- Discuss how knowledge-driven RAG and GraphRAG can make retrieval more trustworthy
- Take you through 3 applied use cases and show the effect of different approaches

Register here: https://t.co/jmfxNa7lO6

[DataTalksClub @DataTalksClub]

How can people understand, evaluate, and appropriately rely on AI advice for high-stakes decisions?

Daan Di Scala researches how structured knowledge can make LLM systems more trustworthy in agrifood, healthcare, and legal.

[DataTalksClub @DataTalksClub]

Daan is an AI researcher at TNO, the Dutch research institute for applied science, and studies AI trustworthiness in high-stakes domains.

Daan's PhD at Utrecht University is on Knowledge-Grounded AI for Trustworthy Advice: how structured knowledge can enhance LLM systems.

[DataTalksClub @DataTalksClub]

That comparison comes from Module 4 of Machine Learning Zoomcamp. The lesson walks through it step by step, with code: https://t.co/14HPfI2sxk

[DataTalksClub @DataTalksClub]

A churn model with 80% accuracy sounds good.

Here comes the uncomfortable check:

A dummy model that predicts no customer will ever churn has 73% accuracy, because only 27% of the customers actually churned.

The improvement over doing nothing is much smaller than you thought.

[DataTalksClub @DataTalksClub]

Here's what can help:

1. Compare against a dummy baseline before reporting accuracy
2. Look at the errors in detail, with a confusion table, precision, and recall

[DataTalksClub @DataTalksClub]

There's a treat too: a $150 code toward the official Airflow 3 certification exam.

Register for Debugging Airflow’s scariest issues:
https://t.co/BB9ezW9p3U

[DataTalksClub @DataTalksClub]

Airflow pipelines have monsters of their own 👻

🧛 Vampire tasks that drain memory from your workers
👻 Ghost tasks that maybe ran, but left no logs in sight

Sound familiar? You're not alone.

[DataTalksClub @DataTalksClub]

Debugging Airflow doesn't have to be scary.

On October 29, Astronomer's Airflow experts will show which context to surface so that you and your AI agent can debug even the spookiest pipeline issues.

[DataTalksClub @DataTalksClub]

This week, Dewang and Suhas are our Book of the Week authors. Ask them your questions about Designing AI Systems in our Slack: https://t.co/SnYQ5rMyP9

4/4

[DataTalksClub @DataTalksClub]

Dewang Sultania went from building early inference platforms at Adobe to deploying multimodal AI models at scale.

In Part 2 of our series on building AI platforms, we talked with Dewang about what it takes to ship a production agent.

1/4

[DataTalksClub @DataTalksClub]

The recording is now up: https://t.co/Ay5TB0C6Nd

Part 1, with Suhas Suresha on SDK design, model routing, vector search, and evals: https://t.co/hYRj1ujlJ1

3/4