Gul Saeed Khattak

Lead Data Engineer at the University of Glasgow, exploring the future of data and AI through autonomous agents, and agentic workflows

[Gul Saeed Khattak @Ldataengineer]

Cloudflare's database D1 now manages millions of databases. One per user, per app, per project.

Here's why that matters, and why it's suddenly a big deal.

D1 is a serverless database built on SQLite. You create a database with one command, pay only for what you use, and never run a server.

The idea isn't new. What's new is the scale.

AI app builders are changing how apps are made. Every "build me an app" prompt can need its own little database. OpenAI's ChatGPT Sites uses D1 for exactly this: leaderboards, comments, metrics.

Live D1 databases grew 15x in a year. 65x in two.

The hard part isn't creating millions of databases. It's managing them without slowing every query down.

Cloudflare's fix, and lessons worth stealing:

→ Keep the hot path clean. Queries never touch the central database
→ Cache routing info close to users
→ Precompute totals ahead of time. Don't run SUM() over millions of rows on every request
→ Use queues for anything that can be a few seconds late, like billing
That third one is the same lesson GitHub just shared about star history: do the heavy counting before the user asks.
One honest note: D1 isn't open source. It's a managed Cloudflare service built on open-source SQLite.
https://t.co/cxtW9ZSW3v

[Gul Saeed Khattak @Ldataengineer]

Gemini 4 Argon, Google’s newest frontier model, is now generally available in GitHub Copilot.

[Gemini 4 Argon in GitHub Copilot - GitHub Changelog](https://t.co/ahAh7O54dN)

[Gul Saeed Khattak @Ldataengineer]

Why would US tech giants warn Europe about the cost of leaving US tech giants?

Follow the money.

Europe is planning a new law, the Cloud and AI Development Act.
It could push governments, and maybe key industries, toward European cloud providers.

Now a new study says that could cost Europe €115 billion over 10 years. That's more than 6x the EU's own top estimate of €18 billion.

Who paid for the study?
CCIA Europe, a lobby group whose members include Amazon, Google and Meta. The companies that would lose business if Europe switches.

And the €115B has a big asterisk.
About €80B of it assumes the rules spread to 25% of cloud work in private sectors like banking and healthcare.
That's a modelled scenario, not a real bill.

There's another irony.
Amazon and Google already sell "sovereign cloud" products in Europe.
Sovereignty is fine, apparently, as long as it has their logo on it.

But here's the part I wouldn't dismiss.

69% of the estimated cost comes from rewriting and re-architecting software.

That's the real lesson, whoever paid for the study. The more your code depends on one provider's special services, the more it costs to leave.

Lock-in isn't a bill you get when you sign up. It's the bill you get when you try to leave.

https://t.co/lK4lmHvfqx

[Gul Saeed Khattak @Ldataengineer]

Palantir vs London's mayor: the £50M AI police deal fight is over
Palantir and London's mayor have dropped their legal fight over a £50M police AI contract.

Quick background:

In May, the mayor's policing office blocked the Met Police from signing a two-year deal with Palantir. The AI system would have automated intelligence analysis in criminal investigations.

The mayor's office said the Met broke procurement rules. Palantir was the only supplier seriously considered, and the Met never got its plan approved.

Palantir sued. It claimed the deal was really blocked because of objections to its "values and ethics."

Now, a settlement:
→ Both sides withdrew their claims
→ The mayor's office says its decision "was not based on Palantir's values or ethics"
→ No damages paid, no admission of fault, each side covers its own legal costs
→ The contract is being re-run, and Palantir can bid again
The bigger question hasn't gone away:
When governments buy AI for policing, how much should a company's reputation matter? Or should it only be about cost, rules and whether the tech works?
Where do you stand?
Link in the background👇

[Gul Saeed Khattak @Ldataengineer]

The AI race is back on. And it's four horses now.
A week ago, the AI race looked like a two-horse race: OpenAI and Anthropic.

Google seemed far behind. Meta had spent billions and still hadn't caught up.

Then, in a few days, everything changed.

→ Google rolled out Gemini 4 Argon worldwide. Strong at coding, long tasks, and finding and fixing security bugs on its own.
→ Meta launched Muse Wonder and Muse Wonder Full. It says they match OpenAI's and Anthropic's best on coding and professional work.
Suddenly it's a four-way race again.

But a few things worth noticing:

→ These are the companies' own benchmark claims. Independent tests will tell the real story.
→ Meta's comeback is closed, not open. The company famous for free Llama models now sells its best AI through paid plans and an API.
→ Winning benchmarks isn't winning users. Meta's Muse app has around 9 million downloads, but growth is slowing and it keeps users less well than ChatGPT.

For developers, this is good news either way. More real competition means better models, lower prices, and less dependence on one or two labs.
The AI race isn't over. It just got its second wind.

[Gul Saeed Khattak @Ldataengineer]

GitHub went from 2,500 requests to 24. The lesson is pure SQL.
Filter first, then count: how GitHub fixed star history

And the lesson behind it is one every data person should know.

The story:

Developers love "star history" charts. They show how a project's GitHub stars grew over time, and whether a launch or release actually worked.

But GitHub had no simple way to get that chart. So tools downloaded the full list of every person who starred a repo, kept the dates, and threw the names away.

Two problems:

→ Spammers used those same lists to collect usernames for spam and phishing
→ For popular repos, it was slow and expensive. Building the chart for React took about 2,500 requests
So GitHub built a new API that returns just daily star counts. No names attached.

Here's the clever part.

Returning less data didn't make it cheaper. Their first idea still made the database read every star, group them, sort them, and only then cut out the page you asked for.

The fix: filter by date first, then count.

Now each request only touches the weeks it returns. React's full history: about 24 requests instead of 2,500.

Two lessons worth stealing:

→ A smaller result doesn't mean a cheaper query. LIMIT doesn't save the work the database already did. Filter early.
→ Only return what people actually need. Anything extra is data someone can misuse.

Better performance and better privacy, from the same change.
https://t.co/WGOlIUNXVv

[Gul Saeed Khattak @Ldataengineer]

https://t.co/EvfvriiJOs

[Gul Saeed Khattak @Ldataengineer]

Someone built a free, open-source alternative to Canva. And you can run it on your own server.

It's called Doory.

It's a full design editor in your browser. Posters, social posts, slides, resumes, thumbnails, business cards.

What stands out:

→ Every feature is free. No Pro plan, no locked templates, no watermarks
→ No account needed to start designing
→ Background removal runs right in your browser
→ Animations, music and video export: MP4, GIF, plus PNG, SVG and print-ready PDF
→ Real-time co-editing with live cursors and comments
→ Self-host it with one docker compose command, so your designs stay on your own server
Honest notes before you switch:
→ It ships with only 10 starter templates. Canva has a huge library. You'll need to build your own
→ It's very new: just 6 commits in the repo so far
→ The licence isn't MIT. You can use it commercially, but you must show a credit to the author inside the app
Still, for schools, small teams, or anyone who doesn't want their designs on someone else's cloud, this is worth a look.

Link to the GitHub Repo in first comment:

[Gul Saeed Khattak @Ldataengineer]

https://t.co/BwotART6YL

[Gul Saeed Khattak @Ldataengineer]

The AI race is back on. And it's four horses now.
A week ago, the AI race looked like a two-horse race: OpenAI and Anthropic.

Google seemed far behind. Meta had spent billions and still hadn't caught up.

Then, in a few days, everything changed.

→ Google rolled out Gemini 4 Argon worldwide. Strong at coding, long tasks, and finding and fixing security bugs on its own.
→ Meta launched Muse Wonder and Muse Wonder Full. It says they match OpenAI's and Anthropic's best on coding and professional work.
Suddenly it's a four-way race again.

But a few things worth noticing:

→ These are the companies' own benchmark claims. Independent tests will tell the real story.
→ Meta's comeback is closed, not open. The company famous for free Llama models now sells its best AI through paid plans and an API.
→ Winning benchmarks isn't winning users. Meta's Muse app has around 9 million downloads, but growth is slowing and it keeps users less well than ChatGPT.

For developers, this is good news either way. More real competition means better models, lower prices, and less dependence on one or two labs.
The AI race isn't over. It just got its second wind.

[Gul Saeed Khattak @Ldataengineer]

GitHub went from 2,500 requests to 24. The lesson is pure SQL.
Filter first, then count: how GitHub fixed star history

And the lesson behind it is one every data person should know.

The story:

Developers love "star history" charts. They show how a project's GitHub stars grew over time, and whether a launch or release actually worked.

But GitHub had no simple way to get that chart. So tools downloaded the full list of every person who starred a repo, kept the dates, and threw the names away.

Two problems:

→ Spammers used those same lists to collect usernames for spam and phishing
→ For popular repos, it was slow and expensive. Building the chart for React took about 2,500 requests
So GitHub built a new API that returns just daily star counts. No names attached.

Here's the clever part.

Returning less data didn't make it cheaper. Their first idea still made the database read every star, group them, sort them, and only then cut out the page you asked for.

The fix: filter by date first, then count.

Now each request only touches the weeks it returns. React's full history: about 24 requests instead of 2,500.

Two lessons worth stealing:

→ A smaller result doesn't mean a cheaper query. LIMIT doesn't save the work the database already did. Filter early.
→ Only return what people actually need. Anything extra is data someone can misuse.

Better performance and better privacy, from the same change.
https://t.co/WGOlIUNXVv

[Gul Saeed Khattak @Ldataengineer]

https://t.co/ms7N39LLOQ

[Gul Saeed Khattak @Ldataengineer]

Palantir vs London's mayor: the £50M AI police deal fight is over
Palantir and London's mayor have dropped their legal fight over a £50M police AI contract.

Quick background:

In May, the mayor's policing office blocked the Met Police from signing a two-year deal with Palantir. The AI system would have automated intelligence analysis in criminal investigations.

The mayor's office said the Met broke procurement rules. Palantir was the only supplier seriously considered, and the Met never got its plan approved.

Palantir sued. It claimed the deal was really blocked because of objections to its "values and ethics."

Now, a settlement:
→ Both sides withdrew their claims
→ The mayor's office says its decision "was not based on Palantir's values or ethics"
→ No damages paid, no admission of fault, each side covers its own legal costs
→ The contract is being re-run, and Palantir can bid again
The bigger question hasn't gone away:
When governments buy AI for policing, how much should a company's reputation matter? Or should it only be about cost, rules and whether the tech works?
Where do you stand?
Link in the background👇

[Gul Saeed Khattak @Ldataengineer]

https://t.co/Ko2L3tm92W

[Gul Saeed Khattak @Ldataengineer]

AIPOCH Wants to Be the VS Code of AI Research

AIPOCH Open-Science v1.0 just dropped.

What caught my attention is that it tries to put the messy parts of research in one place:

Literature → scientific data → AI-assisted analysis → inspectable results → portable research records.

And importantly, it is open source + local-first.

That means your research workflow doesn’t have to live entirely inside someone else’s cloud.

Think of it as an AI research desk where the papers, data, agents and results can actually stay connected.

Runs on Windows, macOS and Linux.

For researchers building with AI, this is the kind of tooling I want to see more of:
not just an AI that gives answers , but a workspace that helps you understand how you got there.

[Gul Saeed Khattak @Ldataengineer]

@NVIDIAAI @ClementDelangue Funny week for NVIDIA. Every model launched by others ran on your GPUs. Now you're launching one yourself. Selling shovels and also digging for gold.

[Gul Saeed Khattak @Ldataengineer]

@TechCrunch HTAP isn't new.
It's 10+ years old.
Agents just gave it a reason to matter

[Gul Saeed Khattak @Ldataengineer]

That explains it.

Your feed has become my go-to for catching new launches before anyone else talks about them.

Thanks for the reply, Robert.

As a data engineer, "lists of 70,000 people" is the most impressive thing I've read today.

That's a hand-built recommendation engine. Loved it !

[Gul Saeed Khattak @Ldataengineer]

AI people: add @Scobleizer to your feed.

He has been tracking Silicon Valley, AI, robotics and emerging tech for years ..and has a habit of spotting interesting things early.

Worth a follow if you want to see where tech might be heading next

[Gul Saeed Khattak @Ldataengineer]

8x growth in AI-agent use by EU small businesses is the real and encouraging signal here.