Things I've written.

A field guide in print and audio, essays on privacy and evals, talks on shipping AI safely. Updated as new pieces ship.

03Book

Now in print

Machine Learning to Gen AI Agents.

A conceptual journey from foundations to practice — eighteen chapters that take a reader from gradient descent to multi-agent systems without losing the thread.

A field guide · Edition I

Machine Learning to
Gen AI Agents

A conceptual journey from foundations to practice.

Prashant Kumar

One narrative arc through the AI you actually have to ship — from classical ML through transformers, foundation models, MLOps, safety, and the agents now arriving in production.

Built for engineers, builders, and curious leaders. Kindle-ready: 18 chapters, appendices, and a 177-term glossary cross-referenced through the book.

Part i

Foundations — gradient descent to SVM, CNN, RNN

Ch. 1–5

Part ii

Sequence & language — LSTM, GRU, attention

Ch. 6–8

Part iii

Transformers — BERT, GPT, foundation models

Ch. 9–11

Part iv

Shipping — MLOps, evals, shadow & canary

Ch. 12–14

Part v

Safety & alignment — red-teaming, NIST AI RMF

Ch. 15–16

Part vi

Gen AI agents — tool use, planning, multi-agent

Ch. 17–18

App.

Glossary — 177 cross-referenced terms

A–Z

Get the book ↗ Listen — sample (audio) Bulk & teaching

05Insights

Recent writing

Essays, briefings, working notebooks.

A lead article, plus two shorter notes. The agent rotates the list as new pieces ship.

Lead essay · 9 min read

The privacy budget every AI product needs.

"We'll add privacy later" is the most expensive sentence in modern AI engineering. Once user prompts have left your perimeter unprotected, there is no later — only an incident.

A privacy budget says: this is the maximum information we are willing to leak per request, this is how we measure it, and this is who gets paged when it's crossed.

The essay walks through how to set one — what to count, what to ignore, and the three failure modes that show up before anyone notices.

Read essay →

Briefing · 5 min

Prompt injection is a supply-chain problem.

Reframing LLM attacks as a software supply-chain question — and what that unlocks for defenders.

Read briefing →

Notebook · live

Evals that actually predict production.

Patterns that have survived contact with real users and real bug reports.

Open notebook →

Talk · 18 min · video

Shipping AI you can defend.

How privacy, security, and product design fit together — recorded at [Conference, 2025].

Watch talk →