Umidjon Z

Most people think AI fails because of the model.

That’s rarely the problem.

After building real RAG systems + agents, here’s what actually breaks:

1. Bad retrieval (the silent killer) You can have GPT-4 level models… but if your retrieval pulls irrelevant chunks → output is...

Most AI demos look impressive

until you try them on real data

That’s where things start breaking: – messy PDFs – tables losing structure – missing context – inconsistent answers

The model isn’t the problem

it’s everything around it

There’s a big gap between: “it works in a...

I’ve been working with OpenClaw recently

set up an agent that: – generates images – turns them into short slide videos – uploads to TikTok

all automatically

What’s interesting is not the content part

it’s the orchestration

the agent is basically: planning → generating → transforming...

I’ve been building a document agent that generates reports from internal data.

You upload PDFs / Excel / docs It pulls the relevant parts Then generates a clean report with sources

The important part wasn’t generation

It was making sure every answer is actually grounded in...