Writing
Topics: a short glossary
Plain-language notes on the ideas that recur in my writing on analytics, AI governance and data, each linked to the essays and code that go deeper. General explanations, not advice.
AI governance · Agentic AI in banking · Multi-agent systems · Goodhart's law · Counter-metrics · Data quality · Churn modelling and decile lift · Credit risk scorecards · DPDP readiness
AI governance
The policies, controls and accountabilities that decide what an AI system may do, who answers for it, and how its decisions can be explained and corrected. In practice it means drawing a line between actions a system can take on its own and actions that need a human checkpoint, a log and an owner.
Read: Faster cars need better brakes · If agents can collaborate, they can trade
Agentic AI in banking
Agentic AI describes systems that plan and take multi-step actions, such as calling tools, querying data or triggering workflows, rather than only answering questions. In banking and other regulated settings the hard questions are about authority: what each agent is permitted to do, spending and approval limits, identity, and an audit trail that a reviewer can follow.
Read: If agents can collaborate, they can trade · Faster cars need better brakes
Multi-agent systems
Architectures in which several specialised agents (search, write, check, plan) hand work to one another. Much of the behaviour, capability and risk lives in the interfaces and checks between agents rather than inside any single model.
Goodhart's law
Often paraphrased as: when a measure becomes a target, it ceases to be a good measure. People, and optimising algorithms, learn to move the number without moving the thing it was meant to reflect.
Read: When the score becomes the goal · Attention is not the same as value
Counter-metrics
Measures paired with a primary target to reveal gaming or unintended side effects, for example complaints alongside handling speed, or repeat contacts alongside closures. Treating metrics as a portfolio, and reviewing samples of real behaviour, keeps a scorecard honest.
Data quality
Whether data is fit for the decision it supports: complete, consistent, correctly defined and traceable. Useful habits include a metric dictionary, automated checks on volumes, nulls and ranges at each pipeline step, named dataset owners and known-answer tests run on every refresh.
Churn modelling and decile lift
Churn models score how likely each customer is to leave. A ranked list is usually judged by lift: how much more concentrated churners are in the top deciles than in a random sample. Comparing a complex model with a simple baseline shows whether the extra complexity earns its keep.
Read: Synthetic churn exploration · churn-propensity-model on GitHub
Credit risk scorecards
Credit scorecards estimate the probability that a borrower defaults. A classic approach bins variables, measures each one's predictive strength with weight of evidence (WoE) and information value (IV), fits a logistic regression and scales the result to points. Performance is commonly checked with AUC, KS, Gini and calibration.
Read: credit-risk-scorecard on GitHub (synthetic data)
DPDP readiness
Preparing for India's Digital Personal Data Protection Act, 2023 and the DPDP Rules, 2025: knowing where personal data sits, the purpose of each processing activity, notices and consent, processor contracts, retention and erasure, breach response and a channel for data principals' rights.
Read: DPDP readiness guide · dpdp-readiness-toolkit on GitHub