Browse by series
Foundations
Core mathematics and concepts behind modern ML — neural-net fundamentals, the softmax + cross-entropy stack, statistical inference. Each note stands alone; together they form the shared language of everything else in Data.
View the series →Hardware & Compute
Where the maths meets the silicon. GPUs, CUDA, Apple's MPS/Metal/MLX stack, and what runs locally vs. needs a rented cloud GPU.
View the series →Transformer Architectures
From the 2017 paper to today's LLMs. Self-attention and the QKV trio, the GPT decoder-only branch, RLHF and alignment, and the modern transformer anatomy you'd actually fine-tune.
View the series →Notes on Reinsurance
A structured lecture series on reinsurance for an audience that knows general insurance and premium accounting but is not deep on reinsurance. Sources: Swiss Re and Munich Re primers, IRDAI regulations, IC-85 courseware, and the published accounts of Indian non-life insurers.
View the series →Every essay
Modern Catastrophe Modelling
Part 10 of Notes on Reinsurance. Where XL prices actually come from — probabilistic event sets, exceedance curves, the four major vendors, and the Indian peril stack. The closing piece in the first arc of the series.
Ind AS 117 and the Reinsurance-Held Module
Part 09 of Notes on Reinsurance. The accounting standard about to displace forty years of treaty-account-based measurement — the three measurement models, CSM on reinsurance held, the loss-recovery component, and what changes for the cedent.
Reading GIC Re's Accounts
Part 08 of Notes on Reinsurance. A walk through the published numbers of India's sole national reinsurer — premium, combined ratio, retrocession, solvency. The numerical view of the institution described in Part 07.
The Indian Reinsurance Market
Part 07 of Notes on Reinsurance. GIC Re, the four pools, the order of preference, and the institutional structure of the residual market — what the regulations look like from the inside.
Treaty Accounting and the Cash Side
Part 06 of Notes on Reinsurance. The quarterly treaty account, premium and loss reserves withheld, profit commission, and the working-capital story that nobody talks about. Carter Ch. 14 + IC-85 Ch. 9 read together.
How Retention Is Actually Set
Part 05 of Notes on Reinsurance. Loss distributions, capital, and the acceptable-fluctuation problem. Beard's ruin-probability framework, the Indian practitioner's rules of thumb, and what a reinsurance committee actually decides.
Reading a Treaty
Part 04 of Notes on Reinsurance. The clauses that actually do the work — a walk through proportional treaty wording, anchored on R.L. Carter's clause-by-clause exegesis. The operative clause, the retention, the commission scale, follow the fortunes.
How India's Private Non-Life Insurers Actually Use Reinsurance
What the published accounts of 15 private general insurers reveal about cession strategy, segment-level retention, captive group channels and the commission economics that now sit at the centre of underwriting returns — a three-year comparative across FY 2022-23 → FY 2024-25.
Form B-RA, Line by Line
Part 03 of Notes on Reinsurance. Reading the reinsurance schedule in a real Indian non-life insurer's published account — New India Assurance FY 2024-25. Cession ratios by line, the reinsurance balance, and what an analyst should walk away with.
What Reinsurance Actually Does
Part 02 of Notes on Reinsurance. A working definition, the four economic functions, and the four ways reinsurance is structured — distilled from the Swiss Re and Munich Re primers, with the Swiss Re numeric examples worked through.
Indian Reinsurance Regulation, Layer by Layer
Part 01 of Notes on Reinsurance. Where the rules actually live — from the Insurance Act 1938 to the FY 2026-27 obligatory cession notification. The three regulatory layers, the current 4% number, and where it bites you in practice.
MPS, Metal, MLX & CUDA
The GPU-compute landscape for machine learning — why GPUs, what CUDA / Metal / MLX each are, unified memory on Apple Silicon, and what runs locally on a Mac versus what needs a cloud GPU.