хамза хан

💼 Software Architect
👔 x comsian
🙍‍♂️ Pathan
⛰ Belong to mountains

[хамза хан @hamzaashes1]

@patrickhannigan No agenda beyond curiosity, would you be open to a short call to talk through how you're thinking about the agent/grounding layer? Happy to share what's worked and hasn't in my experience.

[хамза хан @hamzaashes1]

@patrickhannigan The problem you're describing (agents that act on a marketer's behalf, grounded in real campaign data rather than generic output) is close to things I've built before.

[хамза хан @hamzaashes1]

Every Model You Trust Is Built on the Same Flaw
https://t.co/pLaW0V7cXe

[хамза хан @hamzaashes1]

Check out my latest article: Your AI coding agent is shipping blind https://t.co/Omva8PbbRg via @LinkedIn

[хамза хан @hamzaashes1]

5/ We're accumulating debt in code, people, and knowledge simultaneously. At AI speed.

So genuine question:

What are YOU doing to solve this?

Explanation gates before merging AI code? Intent audits? Pair reviews?

Or shipping fast and hoping for the best?

Reply 👇

[хамза хан @hamzaashes1]

AI writes 30-40% of prod code now.

We track technical debt.

But 3 papers from 2026 say there are two more kinds silently destroying your codebase — and nobody is measuring them.

Thread 🧵

[хамза хан @hamzaashes1]

4/ The paper calls them "Fragile Experts."

High output. Critically low corrective competence.

They look like your best engineers — until production breaks at 2am and the AI is down.

[хамза хан @hamzaashes1]

For founders and engineers watching AI infrastructure costs climb every month, this is the kind of problem worth paying attention
Follow my page if you want updates as this goes live.

#BuildingInPublic #LLM #AIEngineering #RAG #Startups #BackendEngineering

[хамза хан @hamzaashes1]

Most startups building on top of LLMs are quietly burning money on tokens.

Not because their prompts are bad.
Because their retrieval pipeline is inefficient.

[хамза хан @hamzaashes1]

I’m currently implementing it from scratch and benchmarking it against a real production codebase.

Early results are showing 60–70% token reduction without sacrificing output quality.

[хамза хан @hamzaashes1]

Once guest RAM and vCPU state are gone, “resume” is a lie. It’s a replay.

This is why agents feel flaky:

They’re logically continuous but physically rebooted.

[хамза хан @hamzaashes1]

Following up on my last post

A few people may asked: “Why not just checkpoint containers or restart faster?”

Because at the VM level, most platforms don’t actually pause execution.

They terminate the VMM, drop guest RAM and reboot from disk.

[хамза хан @hamzaashes1]

But disk isn’t the agent’s brain.

RAM is.

Things agents keep only in memory:

- vector indexes

- runtime caches

- interpreter & planner state

- browser / JS engine state