Kaimanov

PhD in AI/ML · advisor to startups. LLM systems, RAG and agents that make it to production. Short lessons, real fixes, no hype. For founders and ML teams.

[Kaimanov @ka1manov]

"Learn backpropagation" is a roadmap item you can tick without being able to do it.

In my handbook it reads: implement backprop for a two-layer net in NumPy and match PyTorch's gradients to 1e-6. You either have that artifact or you don't.

It's a free AI/ML engineering handbook for people who ship ML systems, not notebooks. No signup, no email, no tracking.

All 42 roadmap nodes get the same three lines:
- prove it: an artifact, not an activity
- why it matters: what breaks in production without it
- an honest time estimate for someone studying around a job

Also inside:
- a RAG deep dive with a failure-mode taxonomy: symptom → cause → diagnostic → fix
- 122 interview questions and, where it matters, a weak answer next to a strong one
- a startup playbook: build, buy or fine-tune, with the cost arithmetic shown
- 3 calculators: GPU memory and KV cache, cost per successful outcome, vector index sizing
- 50 field notes and 6 long reads

Don't start at stage 00. Read the prove-it lines, and whatever you couldn't produce this afternoon is your list.

Link in the reply.

What would you make an ML engineer prove before you'd trust them with production?

[Kaimanov @ka1manov]

Where the rebuild drifts is the useful part. That gap is usually a filter or a join someone added upstream and never wrote down.

I look for the mirror version in backtests: a strategy that comes out clean on the first run usually has a look-ahead bug in it, not a result.

Do you rebuild from raw events, or from the tables the report already sits on?

[Kaimanov @ka1manov]

"Learn backpropagation" is a roadmap item you can tick without being able to do it.

In my handbook it reads: implement backprop for a two-layer net in NumPy and match PyTorch's gradients to 1e-6. You either have that artifact or you don't.

It's a free AI/ML engineering handbook for people who ship ML systems, not notebooks. No signup, no email, no tracking.

All 42 roadmap nodes get the same three lines:
- prove it: an artifact, not an activity
- why it matters: what breaks in production without it
- an honest time estimate for someone studying around a job

Also inside:
- a RAG deep dive with a failure-mode taxonomy: symptom → cause → diagnostic → fix
- 122 interview questions and, where it matters, a weak answer next to a strong one
- a startup playbook: build, buy or fine-tune, with the cost arithmetic shown
- 3 calculators: GPU memory and KV cache, cost per successful outcome, vector index sizing
- 50 field notes and 6 long reads

Don't start at stage 00. Read the prove-it lines, and whatever you couldn't produce this afternoon is your list.

Link in the reply.

What would you make an ML engineer prove before you'd trust them with production?

[Kaimanov @ka1manov]

@hallrizon You’re welcome, bro. Hope you will pass your interview

[Kaimanov @ka1manov]

@mjinkov Thanks, your words inspire me❤️

[Kaimanov @ka1manov]

@FReza1984 Spot on. Without verifiable outputs or runnable code at each step, AI roadmaps just turn into an endless loop of theoretical rabbit holes.

[Kaimanov @ka1manov]

Source on GitHub, prose CC BY 4.0, code MIT. Found a mistake? Open an issue:
https://t.co/VXsMjFFeS4

Read it here, free, no signup:
https://t.co/nbk9yZ4jtb

[Kaimanov @ka1manov]

@noor7si welcome everyone, with you to grow together, I am happy to see you all

[Kaimanov @ka1manov]

@This_is__Now GM everyone, nice be part of this 💌

[Kaimanov @ka1manov]

@spacefundz Guys hello, if you interested in AI/ML, know more about agents, my profile can be interesting for you

[Kaimanov @ka1manov]

@mayurpatilx Agreed, time will judge if software engineer still alive or not

[Kaimanov @ka1manov]

Why do #coders use AI tools if they think they are better and that #AI can't replace them?

[Kaimanov @ka1manov]

@Shinde_Revati21 That’s such an amazing comment, like it.

[Kaimanov @ka1manov]

Stacking Agentic AI on top of Generative AI makes perfect sense conceptually. But honestly, getting the memory and planning modules to work reliably for autonomous execution is where most production workflows completely break down. You can have the best LLM sitting in that Generative layer, but if it loses context while attempting basic tool use, the whole stack falls apart. Which of these layers usually eats up most of your debugging time?

[Kaimanov @ka1manov]

@hiamitkumar_ 😂 this is how X works, you wanna check latest updates in your account, but start reading something interesting from another guys

[Kaimanov @ka1manov]

@viktoroddy The real test for these 3D models is always the mesh topology and clean UV unwrapping. How well does Codex handle the export into actual production pipelines?

[Kaimanov @ka1manov]

@master_jpma Ship a product. Building in public is the new entry-level internship.

[Kaimanov @ka1manov]

@seftsaint Going from Godot's lightweight simplicity straight into the Unreal behemoth is quite the jump, but the visual fidelity speaks for itself.

[Kaimanov @ka1manov]

@Mf360WW Ohh my gosh 🤯

[Kaimanov @ka1manov]

@Mr_Salio Hello Salio, Could you please share your experience, how did you use deepseek and compere to claude

[Kaimanov @ka1manov]

@PrashantByte09 Don’t try it yet. Do you have the ability to use Geminy for code?
Can you describe your experience?

[Kaimanov @ka1manov]

@hiamitkumar_ Heh, that’s great that now we have a piece of job like this. Soo AI can write much better