XAI/Explainable AI Masterclass (incl. Generative AI)

What you'll learn

---

You will learn how AI explainability actually works, from what separates a merely accurate explanation from a useful one to how to design explainability
You will learn the concrete XAI methods for discriminative models, including feature importance, LIME, SHAP, counterfactuals, gradient-based methods, and more
You will learn how to evaluate generative AI explanations, including chain-of-thought and self-explanation faithfulness, confidence and uncertainty, and more
You will learn the hidden costs and risks of explainability itself: gaming, privacy leaks, lost competitive edge, diluted accountability, picking tools, etc
Requirements

---

You don't need any prior knowledge of explainability specifically (a basic understanding of how AI/ML models work helps, but is NOT required, as we recap the discriminative vs. generative distinction from the ground up);
Description

---

OPEN THE BLACK BOX: THE XAI / EXPLAINABLE AI MASTERCLASS
AI models are making more decisions that affect real people every day: who gets a loan, who gets flagged as high-risk, what a legal researcher recommends, what a chatbot claims is true. And every one of those decisions can, in theory, be "explained." But here's the problem nobody tells you...
Most explanations are plausible. Very few are faithful. A discriminative model can hand you a clean, convincing attribution for a decision... that is entirely wrong. A generative model can produce a fluent, logical-sounding chain of reasoning... that has nothing to do with how it actually reached its answer. If you don't know the difference, you're building trust on sand.
Most courses on this topic pick a side: either they cover classic ML interpretability (SHAP, LIME, feature importance) and ignore generative AI entirely, or they cover LLM explainability in vague, hand-wavy terms with no rigor. This course closes that gap. You'll get both model families, side by side, with the actual mechanics, the actual failure modes, and the actual heuristics for choosing a method.
MODULE WALKTHROUGH
The course is structured in three modules
-Foundations of AI Explainability: what "explainable" really means, the traits and audiences that shape a good explanation, the hidden costs and risks explanations create (gaming, privacy, competitive edge, diluted accountability), designing explainability into a model from the start, and the core discriminative vs. generative divide.
-Discriminative AI Explainability: the full toolkit for classifiers and predictive models, feature importance, local attribution (LIME, SHAP), counterfactuals and recourse, gradient-based methods for images, concept-based methods (TCAV), and global surrogate models.
-Generative AI Explainability: how to evaluate what a generative model tells you about itself, chain-of-thought and self-explanation faithfulness, confidence and contradiction calibration, grounding claims in sources and RAG evidence, reading tool calls and agent trajectories as hard evidence, and mechanistic interpretability.
LET ME TELL YOU... EVERYTHING
Here's the full list of what's covered. No vague chapter titles, the actual granular content. Some people just like to ask, "Hey, what is everything I will learn?". Well, here it is
- You will learn what separates a good, useful explanation from one that is merely accurate;
- You will learn the traits and audiences that determine what an explanation should actually contain;
- You will learn the hidden risks of explanations: gaming the system, privacy leaks, lost competitive edge, and diluted accountability;
- You will learn how to treat explainability as a design decision, picking the simplest model that clears the bar rather than bolting on explanations after the fact;
- You will learn the crucial gap between plausible and faithful explanations, and why this gap hits both discriminative and generative models;
- You will learn how discriminative explainability ("what caused this?") differs fundamentally from generative explainability ("do we trust the story?");
- You will learn permutation importance, drop-column importance, built-in tree-based importance, and SHAP-based global importance for ranking what a model relies on overall;
- You will learn local attribution with LIME and SHAP, and when to reach for the fast, intuitive read versus the consistent, defensible one;
- You will learn counterfactual methods: how to find the smallest input change that flips a model's decision, and the "knobs" of sparsity, actionability, plausibility, and diversity that make a counterfactual usable rather than offensive or infeasible;
- You will learn gradient-based methods like Grad-CAM and integrated gradients for explaining image and high-dimensional model decisions;
- You will learn concept-based methods like TCAV for testing whether human-named concepts actually drove a model's predictions;
- You will learn global approximation with surrogate models, and how to report their fidelity;
- You will learn how to match a discriminative XAI method to the use case, audience, and access level (query-only black-box vs. white-box internals);
- You will learn how correlated features split credit in importance rankings, and why a clean importance chart can still hide a proxy for a protected attribute;
- You will learn the faithfulness gap in reasoning-based explanations: chain-of-thought vs. self-explanations, and why reasoning that steers an answer is fundamentally different from a rationalization built after the fact;
- You will learn why a model's stated confidence is usually uncalibrated (and typically overconfident), and how to calibrate it per output category instead of taking it at face value;
- You will learn how to read uncertainty and contradiction signals, and why consistent conflict resolution is a good sign but never proof of correctness;
- You will learn to ground generative claims in checkable evidence: RAG (Retrieval-Augmented Generation), fact blocks, and external tool evidence, and the vocabulary distinction between a source and a claim;
- You will learn why retrieval is not the same as use, why presence of a citation is not proof it drove the answer, and how one-sided retrieval can present a partial view as the grounded whole;
- You will learn how to read tool calls, action outputs, and errors as hard evidence of what an agent actually did;
- You will learn how to treat an agent's full ordered trajectory of steps as an explanation in itself;
- You will learn the basics of mechanistic interpretability: opening the network to read its real internal computations, and where cost and access Requirements

---

justify (or don't justify) the method;
- You will learn how to layer generative explainability methods by cost, starting with cheap evidence logging and uncertainty flags, moving to faithfulness perturbation tests, and reserving mechanistic interpretability for when the stakes demand it;
- You will learn how to match a failure mode to a method: hallucinated facts to grounding and verification, plausible-but-fake reasoning to faithfulness checks, weird agent behavior to trajectory logs, and confidently wrong answers to calibration;
- You will learn a practical heuristic for deciding how much to explain: who can misuse the explanation, how users might change behavior, who can be hurt by it, and where the sweet spot sits between saying everything and saying nothing;
MY INVITATION TO YOU
Remember that youalways have a 30-day money-back guarantee, so there is no risk for you.
Also, I suggest youmake use of the free preview videos to make sure the course really is a fit. I don't want you to waste your money.
If you think this course is a fit and can take your responsible AI/ML model knowledge to the next level... it would be a pleasure to have you as a student.
See you on the other side!
Who this course is for

---

You're any data scientist or ML engineer who ships models and needs to explain, defend, or audit their decisions;
You're any AI product manager, compliance officer, or risk professional who needs to understand what an AI explanation can and can't prove;
Developers, tech leads, and solutions architects building on generative AI who need to reason about faithfulness, grounding, and trust in model outputs;
Any professional curious about how to open the black box of both discriminative and generative AI models, and design explainability in deliberately;
Homepage

---