Halide Yılmaz AI
54 posts
Governance10 minThe EU AI Act's New Timeline: What the 2026 Omnibus Changed and What Is Actually in Force Today
The EU AI Act was not delayed; part of it was. What Regulation (EU) 2026/1744 actually changed, why Article 50 transparency has applied since 2 August 2026, and how the 2 December 2027 and 2 August 2028 deadlines now work, traced to primary sources.
Governance9 minHigh-Risk AI Systems: What an ML Team Actually Has to Build Before December 2027
The Article 9-15 obligations for high-risk AI systems, translated from legal language into engineering deliverables: a risk register, data governance, Annex IV technical documentation, event logging, an accuracy declaration and a human oversight design.
Governance9 minLabelling AI-Generated Content: Article 50, Machine-Readable Watermarks and C2PA
Article 50 transparency obligations took effect on 2 August 2026 and were not deferred. What machine-readable marking actually means, hard binding versus soft binding, C2PA content credentials, and why a watermark usually dies at the first re-encode.
Governance8 minISO/IEC 42001 for Engineers: What an AI Management System Actually Requires — and Why It Is Not AI Act Compliance
What ISO/IEC 42001 actually requires, how its Annex A controls map onto real ML artefacts, what ISO/IEC 42005 and 42006 added, and why the standard is not harmonised and therefore grants no presumption of conformity under the EU AI Act.
Governance8 minThe NIST AI Risk Management Framework in Practice: Turning Govern-Map-Measure-Manage into Engineering Work
Instead of restating the four functions of the NIST AI RMF, this post turns them into engineering work. Why the MEASURE function is the one most teams leave empty, what the Generative AI Profile added, and why the framework is currently under revision.
Governance9 minAI Impact Assessment: How FRIA, ISO 42005 and a DPIA Differ — and How to Run One
A fundamental rights impact assessment, an ISO/IEC 42005 impact assessment and a data protection impact assessment are three different documents. Who each one binds, when it is required, when it applies, and how to run them together without doing the work three times.
Ethics8 minFairness Auditing in Practice: How to Actually Audit a Model — and Why Most Audits Are Theatre
The operational version of a fairness audit: what to do when you may not hold the protected attribute, how to choose a metric knowing the impossibility results, how large a subgroup must be before the estimate means anything, and how to write the report.
Production8 minAI Incident Reporting and Post-Market Monitoring: The 15-Day, 10-Day and 2-Day Clocks
Serious incident reporting runs on a 15-day clock, 10 days in the event of death and 2 days for a widespread infringement. What logging infrastructure, severity taxonomy and on-call path an ML team needs in order to meet those clocks at all.
Data8 minData Provenance and Training-Data Licensing: Can You Prove Where Your Dataset Came From?
How do you record training-data provenance through a preprocessing pipeline? The sharp rise in robots.txt restrictions, the GPAI training-data summary template, where the IETF work stands, and what the 2025-2026 copyright rulings mean for engineering practice.
Data8 minData Contracts for ML: Stopping the Schema Change That Silently Breaks Your Pipeline
A silent upstream change degrades a model without tripping a single data alert. What a data contract is, which sections ODCS defines, how to enforce one in CI, and why the tooling comparison every ranking page repeats is now out of date on ownership and licensing.
Data8 minDatasheets for Datasets: The Missing Half of the Model Card
A model card documents the model, not the dataset. Which questions a datasheet answers, how it maps onto Annex IV technical documentation, and why this is the one document that cannot be produced retroactively.
Metrics10 minAnnotator Agreement and Evaluator Reliability: Kappa, Alpha, and Validating an LLM Judge
Raw percentage agreement is not reliability. How Cohen kappa, Fleiss kappa and Krippendorff alpha correct for chance, how to read the thresholds, and how to validate an LLM judge the way you validate a human annotator.
Evaluation9 minBenchmark Contamination: Why Leaderboard Scores Are Inflated
Test set leakage is no longer one team's cross-validation mistake but a defect in the whole field's evaluation infrastructure. How contamination and saturation inflate scores, and how to build your own private held-out set.
Evaluation9 minIs That 1 % Real? Confidence Intervals and Bootstrap for Comparing Models
One model beating another by a point is usually noise. How test set sampling error, bootstrap confidence intervals, paired tests and multiple comparison correction tell you whether the gap between two models is real.
Metrics9 minModel Calibration: What Does It Mean When Your Model Says "90 % Confident"?
An uncalibrated score makes threshold selection meaningless and silently invalidates every downstream cost-benefit decision. Reliability diagrams, expected calibration error, Platt scaling, isotonic and temperature scaling explained.
Methods9 minConformal Prediction: Measuring Uncertainty With Distribution-Free Guarantees
Conformal prediction turns any model into prediction sets with a distribution-free, finite-sample coverage guarantee. Split conformal step by step, the nonconformity score, and why marginal coverage is not conditional coverage.
Privacy9 minIs Your Model Anonymous? Memorization, Membership Inference and the EDPB's New Anonymisation Guidelines
Does your model still hold personal data? A protocol that maps the EDPB three-criteria test onto measurable attack surfaces: membership inference, training-data extraction and memorization capacity, with what to report.
Production10 minEmbedding Drift: Monitoring for Drift in the Foundation-Model Era
Why a per-dimension KS test fails on a 768-dimensional embedding, what to measure instead, how an embedding model version bump silently invalidates your reference window, and why drift is not degradation.
Production10 minKnowing Your Model Broke Before the Labels Arrive: Performance Estimation Without Ground Truth
If the labels arrive in ninety days, is your model broken now? How confidence-based performance estimation works, the assumptions it rests on, and why it goes quiet in exactly the scenario people reach for it.
Production10 minFeature Stores and Training-Serving Skew: Why Your Model Behaves Differently in Production
Offline the model scores well, in production it behaves differently. The four concrete sources of training-serving skew, why point-in-time correctness is really leakage prevention, and when a feature store earns its operational cost.
Methods12 minCausal Inference and Uplift Modelling: “Who Will Buy?” Is the Wrong Question
A churn model that perfectly predicts who will leave tells you nothing about whom to call. Uplift modelling, the four segments, confounders versus colliders, what to do when you cannot randomise, and why reading SHAP causally is a mistake.
Privacy10 minMachine Unlearning: Can You Really Make a Model Forget Someone?
Making a model forget a person is nothing like deleting a row from a database. Exact versus approximate unlearning, relearning attacks that bring forgotten behaviour back, the proof problem, and the data architecture that actually works.
Methods10 minTime-Series Foundation Models: Zero-Shot Forecasting Now Beats the Classics — But Does It Beat Yours?
The answer flipped and most published content has not caught up. What fev-bench and GIFT-Eval actually report, why LLM-adapted forecasters are a separate story, and how to run the comparison honestly on your own series.
Turkish NLP10 minTurkish AI Data Quality and Benchmarks: What Actually Exists
The Turkish resource landscape exists mostly as scattered GitHub lists and arXiv papers. What actually exists, what each benchmark measures, what the Cetvel finding says about data quality, and how to build your own evaluation set.
Data3 minData-Centric AI: Improving the Data Instead of the Model
Data-centric AI holds the model fixed and improves the dataset instead. How label consistency, coverage gaps and error analysis pull more accuracy out of the same architecture, explained with practical examples.
Evaluation3 minClass Imbalance: Building Models That Do Not Miss the Rare Event
In fraud, failure or rare-disease problems the positive class is tiny. How resampling, class weights, threshold tuning and the precision-recall curve let you build a model that actually works on imbalanced data.
Data3 minActive Learning: Spending a Labelling Budget on the Examples That Matter
Labelling everything is expensive. Active learning picks the examples a model is least sure about, buying more accuracy from the same budget. Uncertainty sampling, diversity, and the cold-start problem.
Evaluation3 minData Leakage and Cross-Validation: Why Your Test Score Is Lying
A validation score that looks too good is usually a sign of leakage. Temporal leakage, group leakage, scaling mistakes, and how to choose the right cross-validation scheme.
Ethics3 minBias and Fairness in AI: Why Models Discriminate
Bias usually comes from the data and the problem definition rather than the model. Historical bias, representation gaps, fairness metrics, and why those metrics contradict each other.
Methods3 minAnomaly Detection: Finding the Outlier Without Labels
Anomaly detection usually has no labels and very few positives. Statistical methods, Isolation Forest, autoencoders, and how to manage the false-alarm rate through threshold choice.
Methods2 minTime Series Forecasting: Trend, Seasonality and Honest Validation
Time series problems obey different rules from ordinary machine learning. Trend, seasonality, stationarity, baselines and walk-forward validation, explained practically.
Methods3 minRecommender Systems: Collaborative Filtering, Content-Based Methods and Cold Start
Recommenders rest on two ideas: similar users and similar items. Collaborative filtering, content-based methods, hybrids, the cold-start problem and the diversity trade-off.
Privacy2 minPrivacy-Preserving Learning: Federated Learning and Differential Privacy
You can train a model without centralising the data. Federated learning, differential privacy, the privacy budget, and what these methods actually guarantee.
Governance2 minModel Cards and AI Governance: Documenting a Model
A model card records what a model does, where it fails, and what data trained it. A practical look at governance, auditability and risk classification.
Vision7 minCNNs and the Basics of Image Processing: How Does a Machine "See" an Image?
An intuitive guide to CNNs through filters, convolution, pooling and layers: how a machine actually parses and "sees" an image, with a short code sketch.
History8 minFrom RNNs and LSTMs to the Transformer: The Evolution of Sequence Architectures
An intuitive guide to the evolution of sequence architectures: how RNNs work, their limits like vanishing gradients and the inability to parallelize, the gates of the LSTM, and why attention won out with the Transformer.
Fundamentals8 minOverfitting and Regularization: Why Models Memorize Instead of Learn
Why do models memorize? An intuitive guide to overfitting and regularization: dropout, L1/L2, early stopping, and cross-validation explained with everyday examples.
Training8 minGradient Descent and Optimization: How Does a Model "Learn"?
An intuitive walk through gradient descent via the loss function, learning rate, SGD and Adam: how does a model really "learn" by measuring and reducing its error?
Metrics8 minClassification Metrics: When to Use Accuracy, Precision, Recall, F1, and ROC-AUC
Starting from the confusion matrix, we explain accuracy, precision, recall, F1, and ROC-AUC through everyday examples and clarify which metric is the right choice and when.
Learning9 minUnsupervised Learning and Clustering: Finding Patterns in Unlabeled Data
An intuitive guide to unsupervised learning through k-means, hierarchical clustering and PCA: how to extract hidden patterns from unlabeled data and choose the number of clusters.
Data8 minFeature Engineering: Meaningful Signals from Raw Data and the Leakage Traps
An intuitive guide to feature engineering: deriving meaningful features from raw data, scaling, encoding, and the data-leakage traps that make models collapse in production.
Interpretability6 minModel Interpretability (XAI): Why Did a Model Decide That Way?
An intuitive guide to SHAP, LIME, and attention visualization: why a model decides the way it does, and how these explanations strengthen trust and accountability.
Data7 minSynthetic Data Generation: Overcoming Scarcity, Protecting Privacy, and Avoiding Model Collapse
Synthetic data is generated data that imitates the patterns of reality. We explain, with intuitive examples, how it overcomes data scarcity, protects privacy, how LLMs generate it, and the risk of model collapse.
Production8 minModel Monitoring and Drift in Production: Concept Drift, Performance Decay, and Retraining Triggers
A practical guide to data and concept drift in production ML models: detecting performance decay, retraining triggers, and keeping your production ML healthy.
Deep6 minThe Attention Mechanism, Intuitively: Query, Key, Value, Self-Attention, and Why It Was Revolutionary
When a model understands a word, which words does it "look at"? An intuitive guide to Query-Key-Value, self-attention, and the Transformer revolution.
Turkish NLP7 minTokenization and the Challenges of Turkish: Agglutination, Vocabulary Explosion, and Efficient Tokenization
Why does Turkish consume more tokens? Agglutination, vocabulary explosion, BPE and SentencePiece, and how efficient tokenization affects cost and speed.
Fundamentals6 minEmbedding Models and Semantic Similarity: Vectors That Turn Text Into Meaning
An intuitive guide to turning text into vectors, cosine similarity, choosing an embedding model, and multilingual embeddings, with everyday analogies and a short code example.
RAG5 minChunking Strategies for RAG: How Splitting Decides Quality
RAG chunking strategies: fixed size, sentence/paragraph, overlap, semantic chunking, and preserving heading context. How splitting affects quality, explained intuitively.
Retrieval8 minReranking and Hybrid Search: Combining BM25 with Semantic Search
An intuitive guide to hybrid search blending BM25 and semantic search, score fusion with RRF, and reranking with a cross-encoder, with short code examples.
Limits8 minThe Context Window and Long Documents: Token Limits, Lost in the Middle, Long Context vs RAG
What is a context window? Token limits, the "lost in the middle" problem, long context vs RAG, and summarization chains (map-reduce, refine) explained with intuitive analogies.
Agents7 minLLM Agents and Tool Use: Planning, MCP, and Multi-Step Tasks
How do LLM agents call tools, plan, and run multi-step tasks? We explain MCP, the opportunities, and the risks with plain, everyday analogies.
Advanced RAG7 minKnowledge Graph + RAG (GraphRAG): Retrieval That Captures Relationships
GraphRAG: retrieval that combines a knowledge graph with RAG. Connections between entities, multi-hop and global questions, and closing classic RAG's gaps, explained intuitively.
Scaling6 minModel Inference and Scaling: Cutting GPU Costs with Latency, Throughput, Quantization and the KV-Cache
A practical guide to latency, throughput, quantization, KV-cache and batching, with everyday analogies, focused on lowering the cost of GPU inference.
Data5 minData Quality and Labeling: A Model's Real Fuel
Garbage in, garbage out: data cleaning, labeling strategies, synthetic data, and honest evaluation sets that make up a model's real fuel.