Pranav Belhekar

By Pranav Belhekar · Volume 1 out now

Atlas

The AI engineer's field guide. Production patterns for RAG, memory, and agentic systems.

Atlas is the book I wish I'd had when I started shipping agentic systems into production. It's not a survey of the field — it's a working guide, written from the inside of systems that have actually run, broken, and been put back together. Volume 1 covers the first four chapters; Volume 2 is in progress.

Volume 1 · out now

4 chapters

  1. 01

    Foundations: when an LLM is the right tool

    When to reach for a model, when to reach for a rule. Cost and latency budgets before you write a prompt.

  2. 02

    Retrieval that doesn't lie

    Why naive top-k breaks in week three. Chunking strategies that survive a real corpus. Honest evals.

  3. 03

    Tool use and the contract problem

    How agents fail when tool descriptions drift from tool behavior. Patterns for keeping them in sync.

  4. 04

    Production memory systems

    Short-term, long-term, and the awkward layer between them. What survives a context-window reset.

Volume 2 · in progress

4 chapters

  1. 05

    Multi-agent orchestration without deadlocks

    in progress

    Supervisor patterns, message bus designs, the failure modes that only show up under load.

  2. 06

    Evals that actually catch regressions

    Why your golden set isn't enough. Adversarial sets, drift detection, when to ship anyway.

  3. 07

    Cost, latency, and the boring trade-offs

    Caching, batching, model selection. The unsexy work that decides whether the system survives Q2.

  4. 08

    Observability for agentic systems

    Tracing across tools and models. Knowing when production is misbehaving before a user tells you.

This is for you if

  • You're shipping an LLM-backed product and the demo-vs-prod gap is hurting.
  • You can read TypeScript or Python and want code, not analogies.
  • You've already tried "just use LangChain" and it didn't end where you hoped.
  • You care about evals, latency budgets, and cost — not just clever prompts.

It's not for you if

  • You're looking for an intro to what an LLM is.
  • You want vendor comparisons or "best model for X" leaderboards.
  • You want hype, predictions, or anything about AGI timelines.
  • You expected no code.

Get the book

Digital · Gumroad

Atlas — Volume 1

  • All four Volume 1 chapters
  • PDF + web reader
  • Free updates within Volume 1
Buy on Gumroad →

Print · Amazon

Atlas — Volume 1, paperback

  • Same content, in print
  • Ships globally via Amazon
  • For the people who still annotate in pen
Buy on Amazon →

Volume 2 ships chapter-by-chapter. Subscribe to the field notes below and I'll send you a note when each chapter lands.

Get notified when Volume 2 chapters ship

Plus the rest of the field notes. Roughly twice a month.

Start with Volume 1

Four chapters of working patterns from production agentic systems. Digital, paperback, or both.