Hi, I'm Irai Kumaran

IBM Watson Telco, 7,043 customers, 26.5% churn. A degenerate classifier that predicts "stays" for everyone scores 73.5% accuracy while catching zero churners, which tells you immediately that accuracy is the wrong scoreboard.

We wrote the business cost down explicitly: C = 5·FN + 1·FP. Under an asymmetric cost the conventional 0.5 decision threshold is provably suboptimal, and the best blend of classifiers is not the blend that maximises a symmetric score. So rather than training a stacking meta-classifier on a generic objective, we treated combination as a search problem: Simulated Annealing over a five-dimensional space, four simplex-constrained ensemble weights plus the decision threshold, minimising that cost directly. A Bayesian Belief Network supplied a second opinion on borderline customers.

The most accurate, highest-F1 configuration we built was also the most expensive one to deploy. Selecting a model by symmetric metrics on an asymmetric problem isn't merely imprecise. It's actively misleading. That finding is methodological, so it transfers well beyond churn.

Conventional cybercrime analysis flattens demographic context, which limits how precisely prevention campaigns, law-enforcement resourcing, and regulatory policy can be targeted. We built a pipeline that preserves demographic dimensionality at every stage, every query, every graph traversal, every dashboard filter starts from a demographic lens.

A 5,000-record dataset whose probability weights are calibrated against eleven real government and industry reports from nine countries: FBI IC3, the FTC Consumer Sentinel Data Book, AARP fraud research, Pew, Statista and others. Synthetic records, real distributions.

A normalised four-table MySQL star schema enforcing referential integrity across victim, incident, platform and financial dimensions. A Neo4j property graph of 5,040 nodes and 10,000 directed relationships encoding victim–crime–platform trichotomies as traversable edges. And a Power BI dashboard with six analytical views built for non-technical decision-makers. Analysis ran through twelve structured SQL queries and ten Cypher traversals.

These are patterns that aggregate analysis simply cannot see, and they map directly onto where prevention budget should go. The framework is deliberately extensible: swap the synthetic source for live national reporting feeds and the same pipeline runs.

Victimisation research overwhelmingly uses regression and univariate statistics: age predicts phishing susceptibility, income predicts loss magnitude, digital literacy predicts reporting. Informative individually, but blind to the way a specific combination of age, income, crime type and platform produces a risk profile qualitatively unlike any single factor.

Reframe victim profiling as graph clustering. Each victim is a node; edges encode cosine similarity across nine dimensions: age group, gender, income, education, crime type, platform, device, reporting behaviour and financial loss. Then run PSCAN.mr, a parallel MapReduce adaptation of the SCAN structural clustering algorithm, implemented end-to-end in Apache PySpark to spread the work across cores. Parameters: ε = 0.9, σ = 0.5, μ = 10.

Nearly a quarter of victims are outliers: profiles unique enough that awareness campaigns built around modal demographic risk never reach them.

Scaling laws hold that more compute, more data and more parameters improve performance predictably, and that genuinely new capabilities emerge at sufficient scale. That second half is the part doing the heavy lifting in a lot of strategy decks.

The Data Processing Inequality: processing data can never increase the information it contains about the original source. Supported by Shannon's source coding theorem and Kolmogorov complexity. The photocopier analogy carries it, a copier can be enormous, fast and expensive, and still never print more information than was on the page.

Scale can extract structure that was already latent in the data, faint patterns nobody had noticed, which is why it looks like magic. But computation cannot create information that was never there. Claims that scale alone conjures new capability sit in tension with established information theory, and that misunderstanding is one plausible reason so many industry AI projects fail.

Research and argument rather than code. Building a case, not a demo. The course also requires using genAI tools directly and comparing what ChatGPT, Claude and Gemini claim about scaling against what the math actually supports. That comparison is part of the evidence.