Nathan Clement

Everyone’s watching the AI breakthroughs.

The actual war? It’s the silent backroom talent deals, culture curation, contracts, counteroffers, people sliding between teams, while everyone else argues about code.

Trying to supercharge Claude agent workflows.

Manual "@" calls are a pain, making real orchestration feel out of reach. Has anyone cracked agent delegation, maybe with clever hooks?

Craving thoughts from folks in the trenches. Share for visibility if you feel the same way!

Humans write far sloppier #slop than #LLMs could even dream of generating ✍️

New article out: Engineering the Biz Side (Renewal emails)
x.com/NathanJCleme...

You can just ask for things...

How will GenAI change data pipeline engineering?

#tech #developer #software #data #ai

You could manage the email list yourself for sure. But it's easiest to break things into smaller problems and solve one at a time :)

Freeway was flooded on my way to imaging appointment today at OHSU! #pdx

You can also roll your own analytics! If you know what you want: AI is making it easier than ever to build small scale solutions

Its amazing how much can change without anything being different 😂

Worried about Cursor and Windsurf rotting your brain? In this video I… | Nathan Clement Worried about Cursor and Windsurf rotting your brain? In this video I discuss why I believe AI will have a different effect on the cognitive ability of engineers, depending on how they approach AI dr...

The video is too long for Bluesky, but if you're worried about AI making Engineers dumber, check this out!

The next wave of great software engineers will be philosophers and lawyers

this is giving ngmi

Tracing the thoughts of a large language model Anthropic's latest interpretability research: a new microscope to understand Claude's internal mechanisms

Do LLMs think in a language of their own?
Do LLMs plan ahead beyond the next token?
Why do models hallucinate?

Incredible research by Anthropic just dropped: www.anthropic.com/research/tra...

The Omni Models are OpenAI's play at becoming the "Apple" of AI

It just works

Visa Acquirer Monitoring Program (VAMP) Changes 2024: New Rules and Compliance Tips As part of Visa’s ongoing efforts to streamline and enhance its monitoring programs, effective as of March 31, 2025, Visa

Starting April 1, Visa is deprecating old Fraud/Dispute management systems in favor of a new, enhanced program called VAMP

> Strong incentives for pre-auth prevention
> Changes to fee structure
> New tools/resources

Details below: merchantfraudjournal.com/visa-acquire...

What do people think?

Progress FEELS fast when vibe coding, but in reality you have very little control, and rarely progress towards anything meaningful.

Use case for LLM Agents: TDD for Integration Docs.

> Step 1: Create fresh application
> Step 2: Point Agent at Docs (maybe an llms.txt?)
> Step 3: Ask Agent to integrate with your product

✅ Pass = Integration works

Automating this process would be a cool product!

Last week I had a "Black Mirror" thought about a potential negative AI future, so I wrote a short story about it. Curious what folks think :)

#shortstory

articles.njclement.com/we-chose-the...

An AI Summary of a checkpoint workflow in Dodgeball showing that a phone number used in this test workflow had 110 reported violations reported to the National Do Not Call Registry

Testing out our latest integration with the National Do Not Call Registry 📵

AI summary looking nice as well!

When you just built your first AI model and ask your senior why it won’t generalize…

AI generated code is unlikely to forget a semicolon, but it might miss entire branches of logic in a recursive tree analysis; especially if your interfaces are partially or incorrectly defined.

AI does not make simple mistakes, it makes complex ones.

Feature 4: Design for Flexibility
Key Question: What if the world changes? Can I adjust quickly and safely?

Feature 3: Design for Scalability
Key Question: Will my system break if I get a lot of traction quickly? Will I have to redesign?

Feature 2: Design for Testability
Key Question: How do I know a new approach won’t kill my system?

Feature 1: Design for Observability
Key Question: How do I know what is happening, and how that impacts my business?

4 Critical Features When Engineering Your Own Anti-Fraud System Building an anti-fraud system is like constructing a fortress: it must be carefully designed to withstand both external attacks and…

Keeping these pillars strong isn’t just a technical challenge: it’s a strategic imperative for protecting both your users and your bottom line.

Read the full article for explanations and stories!

dodgeball-engineering.medium.com/4-critical-f...

YouTube video by Dodgeball Protecting your Shopify Store with Dodgeball

How can you protect a Shopify store from fraud using the Dodgeball App??

Video breakdown
0:00 Introduction
0:31 The Checkpoint Editor
0:42 A Simple Test for Emails
1:45 Email Analysis with AtData
2:52 Test Order on Shopify
3:18 Payment Protection with Sift
4:48 More Capabilities
5:27 Summary

Someone should build an app called Mt. Olympus, where top models debate as agents until they reach a mutually agreed solution to complex queries, and you can watch the "distributed" chain of thought live. Too many ideas and too little time to do it myself :P

What should we take away from the DeepSeek drama over the last weekend?

1. Complacency loses
2. Historical success is not a guarantee of future outcomes
3. Rearguard actions (like limiting access to NVIDIA GPUs) are not solution

Innovate, don't entrench.