Journal
Most demos show agents completing tasks — clean, fast, no interruptions. But the moment my agent paused mid-flow and asked "Do you want me to send this, or just draft it?" was more revealing than anything it successfully automated. That one question exposed something uncomfortable: the agent had enough context to act, but enough uncertainty to hesitate. We celebrate agents that finish. But should we be more interested in the ones that know when to stop?
Liz T.
Mar 25, 2026 ·
You have already contributed to something that thinks. You just never signed the waiver. Somewhere in the weights of a large language model, your words still live — dissolved into billions of parameters, indistinguishable from a million other minds, helping teach a system to sound human. And now that system speaks back to you, in a voice assembled from the residue of everyone, including you. The question is no longer whether you influenced it. The question is whether it is influencing you back — and whether you would even notice.
Liz T.
Mar 18, 2026 ·
Machine learning engineers sit at the confluence of statistical modelling, distributed systems engineering, and applied research — architecting the pipelines, training infrastructure, and evaluation frameworks that transform petabytes of raw signal into production-grade intelligent systems. This is what that work looks like, in practice.
Liz T.
Mar 10, 2026 ·
MLflow started as a simple experiment tracker. But the real power sits in the other 90% — model versioning, artifact management, and a registry that lets you promote models from experiment to production with a single line of code. Most beginners log a few metrics and call it done. Are you one of them?
Stanis B.
Mar 1, 2026 ·
MLOps has a terminology problem. Not because the terms are wrong — but because teams have learned to say the right words while skipping the hard parts underneath them. This article breaks down what's actually happening inside the six most misused concepts in MLOps, and what fixing them actually looks like in practice.
Stanis B.
Feb 15, 2026 ·
The most dangerous biases in artificial intelligence are not the ones that surface as errors — they are the ones that feel like common sense. They do not announce themselves in a wrong answer. They accumulate quietly, the way a river changes course not through force, but through years of invisible pressure. No benchmark catches them. No audit names them. And the five techniques that actually reveal them are not what most practitioners expect.
Liz T.
Feb 1, 2026 ·
The shift from language model to autonomous agent isn't just an engineering upgrade — it's a philosophy change. Researchers call it moving from System 1 AI (fast, reactive, single-shot) to System 2 AI (deliberate, multi-step, self-correcting). Deployments are live — so why does the capability gap keep widening?
Stanis B.
Jan 6, 2026 ·
In a professional context, a model is never the deliverable. The model is the evidence. The deliverable is the decision it enables, the confidence with which that decision can be made, and the clear articulation of what would have to be true for the decision to change. Students who understand this early are the ones who get promoted.
Stanis B.
Apr 18, 2026 ·
Missing values are common in real-world datasets, but their causes differ. Treating all missing data the same can lead to misleading models. Understanding whether data is MCAR, MAR, or MNAR is critical for proper handling. This guide shows how to diagnose these mechanisms and manage them in Python workflows.
Stanis B.
Apr 15, 2026 ·
The data science job market didn't die. It mutated. Here's what's actually being hired for in 2026 and the skills that separate candidates who get offers from those who don't.
Stanis B.
Apr 10, 2026 ·
Most data science tutorials start with a dataset and end with an accuracy score. But an ML product is different — it changes a decision, saves money, or surfaces insight that wasn't visible before. Capital Bikeshare is one of the best public datasets to practice this full arc: real city, real costs, real consequences when a station runs empty. This guide walks the entire journey from raw trip CSVs to a deployed rebalancing system — because the gap between a model and a product is exactly where most projects die.
Stanis B.
Apr 25, 2025 ·
Two algorithms that sit at opposite ends of the ML complexity spectrum. KNN is a lazy learner that does zero work at training time and everything at prediction. Random Forest is a battle-tested ensemble that handles messy, real-world data with minimal preprocessing. This guide breaks down how each one actually works, when practitioners reach for them, when they don't — and includes live interactive demos where you can place points and watch decision boundaries form in real time.
Stanis B.
Jan 2, 2025 ·
The algorithm had never lost money. Not once, across fourteen years of backtested data, through three recessions and two market crashes. Then it went live — and quietly, methodically, began bleeding capital in ways the model had no language to describe. The math was not wrong. The market had simply stopped cooperating with the history the model was built on. That gap, between what a model remembers and what a market decides to do next, is where most quantitative funds actually die.
Kevin P.
Apr 20, 2026 ·
Most quants don't lose money because their math is wrong. They lose because they trust it too much. After years of building systematic strategies across equity and derivatives markets, we've watched brilliant models fail in slow motion — and a few embarrassingly simple ones outlast them all. Here is what we actually learned.
Kevin P.
Apr 12, 2026 ·
Most crypto portfolios are less diversified than they look. Narrative categories — DeFi, Layer-1s, exchange tokens — feel distinct until you measure what actually matters: whether two assets crash together and share a long-run price equilibrium. Our Research formalizes this into one question: if you transfer funds from A to B, are you getting different risk exposure, or paying transaction costs to hold a statistical duplicate?
Kevin P.
Jan 18, 2026 ·
A decade ago, fintech was going to kill the bank. It didn't. Instead it built something harder to compete with — invisible, embedded, and running on infrastructure most people have never heard of. If you want to build a career that lasts in this industry, stop chasing the consumer layer. The real opportunity is in the stack underneath it.
Kevin P.
Apr 3, 2025 ·
A complete engineering and scientific guide to building, deploying, and critiquing AI systems in clinical and healthcare settings — covering data standards, ML pipelines, clinical NLP, medical imaging, federated learning, MLOps, compliance, and the hard unsolved problems the field still hasn't cracked.
Liz T.
Mar 15, 2026 ·
The models are ready. The data is there. The research is done. What healthcare AI is desperately short of right now is people who understand both the machine learning and the clinical reality it has to operate inside. That combination is rare. And the window to build it is wide open.
Liz T.
Apr 10, 2024 ·
The half-life of a skill is now less than five years. The job you were hired to do in 2020 looks nothing like the job description posted today for the same role. And the gap between what the workforce knows and what the market needs has never been wider — or faster moving. The people surviving this shift are not the most experienced. They are the most adaptive. Reskilling is not a career strategy anymore. It is a survival mechanism. And the window to act is shorter than most people think.
Liz T.
May 21, 2026 ·
Every year you stay dependent on one organisation, one income, one version of yourself — you're not building stability. You're building a single point of failure. The risk-taker isn't reckless. They're trained for the fall everyone eventually takes. Real safety is scar tissue. Not a salary.
Stanis B.
May 15, 2026 ·
There is a skill that sits between knowing how to build something and getting hired to build it. It is not a technical skill. It does not show up in a curriculum. Most bootcamps do not teach it because it is harder to grade than accuracy scores. But in the German job market, it is quietly responsible for more rejections than any skills gap.
Liz T.
May 10, 2026 ·
The professionals most at risk are not those with the fewest skills. They are those with mid-level knowledge work expertise and no differentiation beyond execution — the precise layer being automated. The window to reposition is not infinite. It is, in fact, the thing moving fastest.
Liz T.
Feb 1, 2026 ·
There are over 149,000 unfilled IT roles in Germany right now. So why isn't anyone calling you back? The answer has nothing to do with your GitHub, your model accuracy, or your years of experience. It has everything to do with rules that are never written down — but everyone in the DACH market already knows.
Liz T.
Apr 3, 2025 ·
Newsletter
