Explore AI resources, insights, and upcoming events.
Learn how product teams design, test, and deploy AI products through practical guides, expert insights, and real-world experimentation strategies.
How to move past vibe checks: scaling manual AI testing into systematic evaluation
Vibe checks are the right way to start testing AI — and the wrong way to keep going. The step-by-step path from a few happy-path tests to systematic, automated evaluation.
Madalina Turlea
AI evals for product managers: the complete guide for 2026
The complete, practical guide to AI evals for product managers in 2026 — what an eval is, why it's a PM skill, and how to evaluate AI quality whether you have a live feature or just an idea.
Madalina Turlea
How to measure AI feature ROI (when half of teams can't)
Half of product teams name unclear ROI as their #1 AI challenge, and a third track no AI metrics at all. Here's a measurable framework for the return on a single AI feature.
Madalina Turlea
Why AI features fail: the silent failure problem
Only 7.8% of teams say their shipped AI features delivered measurable impact. Why AI features fail silently, why teams don't notice, and how to catch it before users leave.
Madalina Turlea
Who should own AI features in product teams?
AI defaulted to engineering, but the judgment layer where domain expertise lives is where AI features are won or lost. Who should own AI in product, and why the ownership has to move.
Madalina Turlea
How to know your AI feature actually works in production
Most teams ship AI features and wait for complaints. Here's how product managers can know — not hope — their AI is working in production.
Catalina Turlea
What is AI experimentation, and why do you need it?
One idea. One prompt. Five real cases. Several models. Read every response. That's where AI-native products start. An 8-step playbook for product thinkers running their first experiment.
Madalina Turlea
EU AI Act compliance for HR and recruitment AI: building the audit trail
If your product screens CVs, ranks candidates, or scores performance, the EU AI Act now classifies you as high-risk. Here's the audit trail you have to produce.
Catalina Turlea
EU AI Act compliance for credit scoring and insurance AI: what the audit trail looks like
Credit scoring and life-and-health insurance pricing are high-risk under the EU AI Act. Here's the audit trail providers and deployers must produce.
Catalina TurleaEU AI Act compliance for EdTech: the audit trail for AI in education and assessment
If your product handles admissions, grading, level placement, or remote proctoring, Annex III puts you in scope. Here's the audit trail EdTech vendors must produce.
Catalina TurleaEU AI Act compliance for biometrics: the audit trail for remote identification, categorisation, and emotion recognition
Biometric AI is the highest-scrutiny category in the EU AI Act. Here's the audit trail required for remote identification, categorisation, and emotion recognition.
Catalina TurleaEU AI Act compliance for critical infrastructure: the audit trail for safety-component AI
Safety-component AI in road, water, energy, and digital infrastructure is high-risk under the EU AI Act. Here's the audit trail required.
Catalina TurleaEU AI Act compliance for law-enforcement AI: the audit trail for risk assessment, polygraphs, and profiling
Law-enforcement AI carries the heaviest documentation requirements under the EU AI Act. Here's the audit trail providers and deployers must produce.
Catalina TurleaEU AI Act compliance for migration, asylum, and border-control AI: the audit trail
AI used in visa, asylum, and border-control decisions is high-risk under the EU AI Act. Here's the audit trail the regime demands.
Catalina Turlea
Why 90% survey satisfaction and 40% churn can coexist
Your AI feature scores high on satisfaction surveys but users keep leaving. Here's why silent AI failures hide behind good survey numbers.
Madalina Turlea
Your AI always returns an answer. That's why you can't trust it.
AI features never throw errors — they just quietly get things wrong. Here's how product managers catch silent failures before users lose trust.
Madalina Turlea
What thumbs up/down feedback actually tells you about your AI feature
Thumbs up/down feedback on AI features measures user sentiment, not output quality. Learn what binary feedback misses and what to measure instead.
Madalina Turlea
The silent failure loop: how you lose users without a single complaint
AI features don't fail loudly. They fail silently — and users leave before you notice. Here's how product teams break the loop.
Madalina Turlea
Why shipping a mediocre AI feature is worse than shipping nothing
A mediocre AI feature doesn't just underperform — it erodes trust, trains users to ignore you, and poisons every AI feature you ship next.
Madalina Turlea
What does "the AI is working" actually mean?
Most product teams can't define what 'working' means for their AI feature. Here's the definition that separates shipping with confidence from Ship and Pray.
Madalina Turlea
Should you still write PRDs when building AI features?
The programming language is plain English. The prompt IS the spec, so why is the PM three handoffs away? Why "evals are the new PRDs" makes things worse, and what PM-owned AI development actually looks like.
Madalina Turlea
Why your AI evaluation is lying to you
Why most teams automate AI evaluation before understanding what they're evaluating — and the step-by-step approach that actually works.
Madalina TurleaEvals in CI are only as smart as the engineer who wrote the assertion
Putting evals in CI is good engineering. It is also silent on every failure mode the person who wrote the test did not know to worry about — which in fintech, healthtech, and insurance is the ones that ship the bug.
Catalina TurleaIf the harness is the moat, evaluation is how you defend it
If models are becoming interchangeable and the agent harness is the moat, evaluation is what turns the loop into an advantage you can prove — on your domain's definition of right, all the way through the deep runs where the interesting failures hide.
Catalina Turlea
The death of the prompt box: what A16Z's 2026 prediction means for your AI features
What A16Z's 2026 Prediction Means for Your AI Features.
Madalina Turlea
Using AI at scale is not the same as using ChatGPT
Using AI in a chat box is solving one problem. Building AI into a product is solving one problem many times — with the same instructions, for inputs you do not control.
Madalina Turlea
The model selection blind spot: why the newest model is not always the best
Teams default to the newest model before checking whether it fits the task. Frontier models do not win every time — and accuracy, cost, and latency rarely peak in the same place.
Madalina Turlea
AI always returns something, and that is the problem
AI never fails completely — it just returns plausible answers, sometimes wrong. Why ship-and-hope traps teams, and what real iteration on a prompt actually looks like.
Madalina Turlea
Why AI features should not be owned by engineering alone
Engineers build the technical part. Domain experts hold the context that makes AI intelligent. Why prompt and evaluation work belongs with the people who understand the problem.
Madalina Turlea
We tested the viral prompt tricks. Most of them do nothing.
Threatening the model, all caps, high-stakes framing — we ran the viral prompt techniques across nine models on real extraction tasks. Structure and clarity beat hacks every time.
Madalina TurleaMarkdown, XML, or JSON: how to format a prompt so the model understands it
The format you put your prompt in changes accuracy more than you would expect. When to use Markdown, when to reach for XML tags, and when JSON is the right tool.
Madalina TurleaWhen to use RAG instead of putting the whole document in the prompt
Putting the whole policy in the prompt works — until you have many policies or one that has grown too big. The simple rule for when to switch to RAG, and the cost reason underneath it.
Madalina TurleaTokens, explained: what you are actually paying for when you build with LLMs
Tokens are the thing you pay for and the only number you can rely on. What a token actually is, why the same prompt costs different amounts across models, and the hidden reasoning tokens.
Madalina Turlea
Temperature, max tokens, and streaming: the LLM settings that quietly change your output
The API parameters you send with every prompt change how the model behaves. What temperature, max tokens, top-P, reasoning settings, and streaming actually do — and when they matter.
Madalina Turlea
Why ChatGPT gets dumber the longer you talk to it
Long chats degrade for a concrete reason: the context window. The model is not remembering — it is being re-sent the whole conversation every turn, and there is a hard limit.
Madalina TurleaDoes the AI know what day it is? Cutoff dates, web search, and why models differ
Models do not know the current date and stop knowing the world at their cutoff. What the cutoff is, why web search is expensive, and the three dimensions that actually make models different.
Madalina Turlea
LLM-as-a-judge: how to evaluate AI features without checking every answer by hand
Most teams build the wrong judge — vague rubrics, one-to-five scoring, no validation. The three-stage evaluation ladder, why a judge should be binary and single-error, and how one prompt went from 50% to 93% agreement with humans.
Madalina Turlea
Error analysis for AI: turning messy review notes into a fix list
You cannot improve an AI feature until you know how it actually fails — not how you think it fails. The unglamorous front end of every reliable AI feature: read, annotate, group by root cause.
Madalina TurleaBuilding an expense-policy AI agent: what we learned reverse-engineering Ramp
Three prompts, six models, 126 runs. What we learned trying to reverse-engineer Ramp's expense-policy AI agent — including why a more detailed prompt actually made GPT-5 worse.
Madalina TurleaHow to run an AI experiment, step by step
From idea to evaluation: how to write the first prompt, build realistic test cases, pick the right mix of models, review the answers side by side, and turn your notes into something you can scale.
Madalina TurleaHow to build a test dataset for an AI feature
A good test set is the difference between a feature that looks fine in a quick check and one you can actually trust. How to replicate real usage, design inputs that try to break the feature, and grow the set as you learn.
Madalina TurleaStart with the prompt, not RAG: giving an AI feature access to your knowledge
The instinct is to reach for RAG. The advice that works is the opposite: start by stuffing the important knowledge into the prompt, and only add complexity when the volume actually demands it.
Madalina TurleaDeterministic metrics: automating the checks you would otherwise do by hand
Small pieces of code that run on every answer and check for something specific — length, format, language, JSON schema, expected items. How metrics measure (not enforce) and connect back to error analysis.
Madalina Turlea
Infinite personalization without the AI slop
AI lets you personalize content infinitely — and slip into generic slop just as easily. The two failure modes to watch for, and why feeding the model your real data is the difference between insight and filler.
Madalina Turlea
Lessons from one year of AI product building
Key insights from building AI products over the past year.
Madalina TurleaShip AI features with confidence for PMs
Course on shipping AI features with confidence for product managers.
Catalina Turlea
The expert test: how to identify high-value AI features for your product
How to identify which AI features will deliver the most value.
Madalina Turlea
Why ship and learn just does not work for AI features
Why the traditional 'ship and learn' approach fails for AI features.
Madalina TurleaReverse-engineering AI products: from system prompts to cost
Let's look at some of the most popular AI products and their estimated AI cost.
Catalina TurleaMyth busters edition - prompting techniques
Let's test out some popular beliefs about prompting AI.
Catalina TurleaDemystify popular AI features with us - part 1
Breaking down how popular AI features work under the hood.
Catalina Turlea
Personalised activation emails with AI
Live workshop on building AI-powered personalised activation emails.
Catalina Turlea
Personalization demo for Airbnb AI product feature
A demo of AI experimentation for personalisation features.
Catalina TurleaTest an AI feature with us - implement Linear's AI Slack feature
Live walkthrough implementing Linear's AI Slack feature.
Catalina TurleaAI experimentation masterclass: personalised Airbnb description with AI
Your first AI experiment - a live walkthrough for PMs.
Catalina Turlea3 mistakes that make your AI feature a silent churn machine
Most AI features don't fail loudly. They quietly underperform while users quietly leave. Here are the three mistakes driving silent churn.
Madalina TurleaThe 4 most expensive AI evaluation mistakes (and the tender that died)
Four AI evaluation mistakes that cost product teams real money — from blown model selection to a million-euro tender lost to silent failures.
Madalina Turlea3 mistakes PMs make when they let engineers own the AI prompt
Engineers build infrastructure. Domain experts build quality. Here are three costly mistakes that happen when PMs hand prompt ownership to engineering.
Madalina Turlea5 AI testing mistakes that will cost you under the EU AI Act
The EU AI Act demands documentation, traceability, and risk management. Most product teams are still shipping AI on vibes. Here are five mistakes to fix now.
Madalina Turlea3 model selection mistakes that are quietly burning your AI budget
Most teams pick their AI model on reputation, not data. Here are three model selection mistakes silently inflating your costs — and how to fix them.
Madalina Turlea4 mistakes that turn a mediocre AI feature into a user trust killer
Most AI features don't fail loudly. They quietly erode trust. Here are four mistakes product teams make that turn mediocre AI into a reason users leave.
Madalina TurleaThe 3 prompt testing mistakes most teams don't know they're making
Most teams test prompts wrong without realizing it. Here are three specific mistakes that silently kill AI feature quality — and how to fix each one.
Madalina Turlea5 mistakes product teams make when shipping AI features
Product teams keep making the same five AI shipping mistakes. Here's what they are, what they cost, and how to fix each one before users find out.
Madalina TurleaShip AI products that actually work
The email course that takes you from first idea to validated AI — no data science team required.