genai/ fieldnotes
Subscribe
ANAND TOSHNIWAL / GENAI SYSTEMS ENGINEERING

Build intuition.
Debug decisions.
Ship better AI.

Retrieval, agents, evaluation, fine-tuning, cost.
Real production problems, thought through properly,
one scenario at a time.

Explore the scenarios Read it or play with it
Read the incident. Run the experiment. Understand the trade-off.
THINK IN SYSTEMS01 / 06
Retrieve→Rerank→Generate

RAG & retrieval. Finding the evidence is the hard part. The model only words it.

THINK IN SYSTEMSRAG & retrieval✳Agents & tool use✳Evaluation✳Fine-tuning✳Latency & cost✳Safety & guardrails
01 / THE FIELD GUIDE

Real problems.
Decisions that matter.

No walls of theory. Start with a situation,
explore the trade-offs, then make your call.

02 / TWO WAYS IN

Read it, or
play with it.

Same fieldnote, two doors. Take the written guide if you think in sentences.
Take the widget if you would rather move something and watch
the system answer back.

DOOR ONE · THE GUIDE11 MIN READ

For the ones who read.

The whole story in order: foundations, implementation, the architecture decisions and why they went that way.

  • A complete guide, start to finish
  • Seven interactive figures and diagrams
  • A glossary and an exercise
  • Architecture decisions, and why they went that way
Open the reading room
DOOR TWO · THE WIDGETRECORDED RUN

For the ones who poke.

Page 42 of 200 just changed. Illustrative model: 4 chunks per page.Take the controls
THE FIELDNOTEDesigning incremental document updates
03 / BEHIND THE NOTEBOOK

Learning in public.
Building with intention.

Anand

Hi, I’m Anand Toshniwal.

This is my growing collection of GenAI architecture scenarios, from retrieval today to agents, evaluation, fine-tuning and cost as the notebook grows: the questions worth asking, the trade-offs worth making, and the experiments that make them click.

Part portfolio. Part field guide. Always a work in progress.

Start with my first fieldnote

New scenario, straight to you.

A note when a new fieldnote goes up: the problem, the approaches, and the lab. Nothing else, and one click to stop.