How AI actually
works.

13 years building mobile and backend systems. Now I take AI systems apart, one failure at a time, in 2-minute reels.

Watch on Instagram Open the vault

No hype. No tool lists. @ai.transition

next workshop · Sat 14 Nov

Where AI systems break

Five hands-on labs for engineers building with LLMs. Online or Bangalore, from ₹3,000.

See the workshop →

latest reels

One failure. Two minutes.

Each reel takes one way an AI system breaks in production, like a refund paid twice or a chatbot that takes 14 seconds, and draws the mechanism until the fix is obvious.

Part 20

Part 19

Reel cover: Thinking off. Flight missed.

Part 18

Part 17

Part 16

open vault

Every reel. All open.

Each reel has a folder in a public GitHub repo: the whiteboard, the rules that fix the problem, and code you can run.

Half the prompt. Same wait.How would you take an AI chatbot from 14 seconds to 1?

Filter on. Answers gone.How do you build an AI search agent that respects user roles and still finds the right document?

Thinking off. Flight missed.How do you make an agent smarter without training a new model?

Saved tokens. Wrong city.Should an agent remember with a running summary or a memory graph?

Paid twice. Key and all.How do you stop a refund agent paying twice when it already has an idempotency key?

$33,000 to say yesHow do you cut an LLM yes/no check from $33,000 to $160 per million?

All 20 parts in the vault →

about

Abhishek Shukla.

I started building mobile apps in January 2013 and spent the next decade on mobile and backend systems at product companies in Bangalore. Today I'm an engineering manager.

Over the last year I moved into AI systems, and I learn in public. Before a part is scripted, I run the experiment behind it myself, so every "I" in a reel is something I actually did. The workshops are the hands-on version of that, for engineers making the same move.

Abhishek standing on a rock at the top of a hike, ocean and city below — Top of Lion's Head, Cape Town