Towards Data Science

For his latest exploration of agentic AI workflows, Benjamin Nweke zooms in on what happens when a tool breaks — but the dashboard stays green.

"The models reason fine, but [...] the underlying enterprise data isn't accessible, clean, or structured enough for them to act on reliably." Partha Sarkar walks us through six advanced architectural patterns for data practitioners.

Not sure how to add guardrails to your LLM agents, or even whether you should? Shuai Guo unpacks the benefits, and outlines potential approaches to making your agentic systems more resilient.

Learn all about the new profiler included in Python 3.15 — @taupirho.bsky.social presents a detailed, hands-on guide to the profiling.sampling module, known as Tachyon.

Even if neural networks won't settle the measurement problem at the center of Schrödinger’s cat, Samit Ganguly suggests they're already changing the way we see, predict, and protect quantum superpositions.

How many independent answers can you actually get when you ask several models the same question? Utkarsh Mangal explains why the answer is often a smaller number than you'd expect.

What does effective dimension reveal about model redundancy, generalization, and trainability that parameter count, loss, and accuracy do not? Adrian Panasiewicz and Jacob Cybulski launch a new series unpacking the factors that shape model outputs and performance.

Taking your first steps in data science? If you could use an accessible, detailed explainer on skewness and kurtosis, we recently shared Mansi Bramta's comprehensive guide.

"Robust estimators address this problem by limiting the influence of observations that are inconsistent with the fitted relationship." Messy data? Don't let it harm the accuracy of your curves — Aamir Hussain Chughtai explains how to address this common challenge.

In a new deep dive from his Robust Estimation Series, Aamir Hussain Chughtai explains the nitty-gritty details of building accurate curves with polynomials and splines — even when your data is messy.

For her debut TDS article, Mansi Bramta presents a patient, detailed introduction to two key statistics concepts: skewness and kurtosis.

We're thrilled to welcome coauthors Adrian Panasiewicz and Jacob Cybulski to TDS! If you're interested in model efficiency and the inner workings of ML systems more broadly, you're in for a treat.

"The plan is to look at what a vote promises, simulate how it can fail, put a formula on the failure, and then measure it on six real models." Utkarsh Mangal digs into the math of correlated mistakes.

What if you could turn a small open LLM into a fast text classifier? Anubhab Banerjee shows how to build a JEV model from an open-source Qwen LLM by replacing its language-modeling head.

How can a machine learn which action produces the best outcome through trial and error? Carolina Bento explores this core idea of reinforcement learning with a practical multi-armed bandit simulation in Python.

LLMs or decision models? Why not both! Shuai Guo's hands-on guide to hybrid LLM workflows shows how to build browser automation by combining decision models with LLMs.

You'd think that adding an agentic element to RAG would consistently improve results. As Abdullahi Dattijo demonstrates, the reality is more complicated than that.

Raw Raman spectra are rarely ready for modeling. Nicolas Coca Lopez walks us through the full pipeline (cropping, spike removal, baseline correction, normalization), and uses non-negative least squares to track a zeolite phase transformation.

"If the hint points at a wrong option and the model moves to it, there is a good chance the hint did that." Utkarsh Mangal digs into the inner workings of LLMs' reasoning process by planting wrong hints and recording the models' responses to them.

Can a model teach itself new skills without forgetting old ones? Jonathan Yahav breaks down on-policy self-distillation, the technique behind dozens of papers since January, and explains why it seems to reduce catastrophic forgetting.

We're thrilled to welcome Anand Rawat to TDS — we invite everyone to explore his excellent article on transforming long-running Python simulations into a self-service platform.

Don't miss the latest installment in Ibrahim Salami's data engineering learning journey — this time, walking us through the process of building an incremental pipeline with dbt snapshots.

For an incisive look at the areas where AI models hit a wall, read Lars Benedikt Kaesberg and Jan Philip Wahle's debut TDS article, which focuses on spatial reasoning and the grid puzzles that trip up powerful language and vision models.

AI can do more than speed up everyday tasks; it can also help us learn more effectively. Destin Gong shares a practical framework for using AI as a thinking partner, from finding resources to practicing active recall.

As AI agents move from individual tools to coordinated fleets, governance becomes harder to overlook. Amber Roberts explores how organizations can protect data, manage access, and deploy agents responsibly at scale.

The challenge with AI is no longer just getting it to do the work; it is finding the time to check the results. Gursimar Singh explains how this shift changes the way we think about assigning tasks to AI agents.

What changes once an LLM comes with a million-token output window? Benjamin Nweke walks us through the nitty-gritty details of using Google's new Gemini 4 Argon model.

For his debut TDS article, Devashish Thakar presents an accessible deep dive on conformal prediction and its practical use cases.

Writing at the intersection of machine learning, computer vision, and medical AI, Renish Joshi shares a follow-up article zooming in on sensor-reduction testing, feature ranking, and other pragmatic factors to consider when working with electroencephalograms (EEG).

We're thrilled to welcome Renish Joshi to TDS with an incisive debut article, focusing on the challenges of participant-aware evaluation in the context of EEG models.

In a thorough and engaging walkthrough of a recent computer vision project, @trflorian.bsky.social unpacks the process behind synthetic-to-real transfer leveraging frozen DINOv3 features.

"A better model follows advice more often but not always. Setting the temperature to zero makes its answers more repeatable, not more obedient." In the latest installment of his Towards Spec-Driven Test Automation series, Gal Arav weighs the benefits and drawbacks of using skills vs. an MCP server.

Switching to a cheaper model should... lower your overall costs, right? Nhu Hoang explains why that's not always the case, and stresses the importance of verifying the hidden costs of LLM routing.

Learn how to turn scattered requiremenets into deployable agents by following along Doster Esh's clear, actionable guide.

"So, let’s talk a bit about the allegations against Character.AI, the technical reasons for them, and how they could have been mitigated." Stephanie Kirmer invites us to examine the technical choices that can lead to AI-related safety failures.

For a fresh, Bayesian look at the Monty Hall problem, don't miss Samit Ganguly's accessible explainer.

Can two charts be accurate and nonetheless contradict each other? The short answer is yes. The longer answer is what you'll find in Sara A. Metwalli's excellent exploration of data storytelling (and its pitfalls).

"The claim, then, is not that the industry's gate is wrong. It is that the gate is structurally incapable of seeing a specific, measurable failure, and that seeing it costs two numbers and an afternoon of measurement." Maxim Podstrechnyy zooms in on the number your quantization gate doesn't have.

A small open-source LLM can do more than generate text. Anubhab Banerjee breaks down how to transform Qwen into a lightweight JEV classifier designed for fast, single-pass decisions.

From early research in the 1950s to modern AI systems, reinforcement learning has a long history. Carolina Bento breaks down the basics using the multi-armed bandit problem as a practical starting point.

What if you used AI to strengthen your thinking instead of simply finding answers? Destin Gong outlines three approaches to building an AI-assisted learning workflow that makes learning new topics more effective.

What changes when organizations move from governing one AI agent to managing an entire fleet? Amber Roberts examines the governance practices needed to keep increasingly autonomous systems secure and under control.

The old AI workflow was simple: ask a question, get an answer, and move on. Gursimar Singh examines why productive AI agents demand a different approach, where delegation and review matter more than writing the perfect prompt.

In a thoughtful new explainer, Utkarsh Mangal explains why LLMs can't perform deterministically even when their temperature is set to 0, and explores why identical completions fall apart after about a hundred tokens.

"Most attacks happen when AI agents are exposed to the lethal trifecta. If agents can read from untrusted sources, access internal knowledge, and communicate to the outside world, they are vulnerable." Thuwarakesh Murallie unpacks the areas where you should add guardrails to your agentic systems.

From a standard ReAct (Reason + Act) loop to autonomous swarms, Partha Sarkar digs into the nitty-gritty details of working with advanced agentic architectural patterns.

Is there *already* a new, powerful rival to Jev? @taupirho.bsky.social takes a close look at OpenAI's new Decisions API and examines its performance on several potential use cases.

How well do AI image judges perform, and what are the areas where they still struggle? Abdullahi Dattijo explores a complex set of challenges in his new deep dive on vision AI models.

"Ten years later I design data architectures for agentic AI systems in production, and recently I re-read our patent with today's eyes. The experience was uncanny." Mario Noioso revisits his 2016 IBM patent and reflects on its similarities to today's AI agents.

"On paper, your name is on the commit. In practice, you only own that code if you can trace its logic and patch it completely unaided." Emmimal P Alexander digs into an issue many practitioners face these days: their shaky sense of ownership over the code they had AI generate for them.