I want everything local – Building my offline AI workspace (instavm.io)

For my sins, I am an EdTech founder building an AI grading platform. We're increasingly finding European customers asking for bundles of documentation to prove safety/efficacy of our platform.

Looking in to the EU AI Act as a few customers have mentioned it and the burden is collosal. We're a team of 10 and don't have capacity to shift to bureaucracy. Genuinely considering avoiding european sales if this becomes a hard requirement.

Anyone have experience navigating the new compliance reqs / know good platforms for compliance automation?

Comments (11)

binarymax · 5h ago

Can you blame them? To me, an AI grading platform for children seems like it should be regulated.

BenJacob1 · 5h ago

My gripe isn’t “no regulation,” it’s the shape of this one for teams in a sector that’s already stretched. Some of these requirements are so high burden that it will add 10% to our OpEx e.g. data governance/provenance reqs: document sources, bias reviews, copyright posture, retention + redoing this after every model change; quality management system across the company (policies, roles, training, audits); conformity assessment + CE marking, with a third-party review all of which will be costly and slow.

I’m all for protecting students. But in practice this shifts months of team time and serious cash into paperwork before schools even pilot the tool. Big incumbents can eat that but small teams either raise prices, slow down, or skip the EU. That’s the innovation tax I’m worried about especially when schools need better tools now.

binarymax · 5h ago

Sure. This seems a cultural difference between US and EU. I've lived extensively in both. EU is known to be burdensome in terms of tech regulations. Normally I find this egregious. However when children are involved I am far more tolerant of their requirements. 10% OpEx overhead to make sure children are protected seems entirely reasonable to me. If you're not prepared to meet their requirements then skip the EU.

quantified · 4h ago

Seconding this. If you're just a team of 10, you are way too small to take on the responsibility of the thing you're trying to provide. I shudder to think of the educational quality a team of ten will provide to hundreds of thousands of schoolchildren in thousands of different schools.

You're too small for the less-regulated US. Add people to deal with things.

BenJacob1 · 4h ago

You're presenting a false dichotomy between building good products and being capable of fulfiling extremely burdensome (not to mention, expensive) compliance requirements from the EU.

We're perfectly qualified to build the core tech. We're just not paperwork experts. Are you suggesting if we can't do paperwork, we shouldn't build innovative products that improve educational outcomes?

quantified · 3h ago

Absolutely. You can build an innovative product, sure, but you can't refine it or support it the way it needs to be supported.

Perhaps you have other teams working on product management, customer success (odd term for education, but applicable), etc.? Are the grading rubrics tuned for each student, what they need, what's in their IEP or 509? What is your plan for handling the legal challenges when the system massively screws up?

This isn't a random B2B. This is a subtly very important influence on a great many actual humans.

magicalhippo · 2h ago

> That’s the innovation tax I’m worried about especially when schools need better tools now.

But how do you, and more importantly the school, know that the tool is better? If you skip the bias review, how can you, and more importantly the school, be sure your AI tool isn't grading every kid with a certain geographical slang less than others for example?

To me this isn't unlike the security stuff, like ISO27001. We make B2B software and had about 15 employees when our customers started pushing ISO27001 or similar hard.

We had no choice but to step up. We have hired multiple people just to deal with that aspect, as well as implementing large changes in our IT infra.

And frankly I think it was a very good thing and has made our org much better and robust.

lordkrandel · 5h ago

Yes, AI deals with sensitive informations, and no one has an idea what it stores about it or what will be done with it. Heavy regulations here. Barriers to making profit by exploiting people and their data. That's the EU. A frontier of freedom and freedom costs. You can have your business in the non-free US where you are a puppet of your tech bro guys. Leave what you don't understand alone.

I want everything local – Building my offline AI workspace (instavm.io)

Ultrathin business card runs a fluid simulation (github.com)

GPU-Rich Labs Have Won: What's Left for the Rest of Us Is Distillation (inference.net)

Tor: How a military project became a lifeline for privacy (thereader.mitpress.mit.edu)

Build durable workflows with Postgres (dbos.dev)

Ask HN: How can ChatGPT serve 700M users when I can't run one GPT-4 locally?

Astronomy Photographer of the Year 2025 shortlist (rmg.co.uk)

Json2dir: a JSON-to-directory converter, a fast alternative to home-manager (github.com)

How we replaced Elasticsearch and MongoDB with Rust and RocksDB (radar.com)

Apollo 13 Commander Jim Lovell has passed away (nasa.gov)

Apple's history is hiding in a Mac font (spacebar.news)

HRT's Python fork: Leveraging PEP 690 for faster imports (hudsonrivertrading.com)

Getting good results from Claude code (dzombak.com)

Linear sent me down a local-first rabbit hole (bytemash.net)

Window Activation (blog.broulik.de)

Fire hazard of WHY2025 badge due to 18650 Li-Ion cells (wiki.why2025.org)

The surprise deprecation of GPT-4o for ChatGPT consumers (simonwillison.net)

Overengineering my homelab so I don't pay cloud providers (ergaster.org)

Telefon Hírmondó (en.wikipedia.org)

A robust, open-source framework for Spiking Neural Networks on low-end FPGAs (arxiv.org)

Voice Controlled Swarms (jasonfantl.com)

Open SWE: An open-source asynchronous coding agent (blog.langchain.com)

Someone keeps stealing, flying, fixing and returning this man's 1958 Cessna (latimes.com)

My ancestors fought in WWII. Hiroshima is plagued by shallow reading (washingtonpost.com)

How attention sinks keep language models stable (hanlab.mit.edu)

Show HN: Trayce – Burp Suite for developers (trayce.dev)

M5 MacBook Pro No Longer Coming in 2025 (macrumors.com)

Job growth has slowed sharply; the question is why (stayathomemacro.substack.com)

Show HN: Synchrotron, a real-time DSP engine in pure Python (synchrotron.thatother.dev)

AI must RTFM: Why tech writers are becoming context curators (passo.uno)

FLUX.1-Krea and the Rise of Opinionated Models (dbreunig.com)

How The Black Cauldron became a notorious Disney flop (bbc.com)

Show HN: Aha Domain Search (ahadomainsearch.com)

Google's Genie is more impressive than GPT5 (theahura.substack.com)

Claude Code IDE integration for Emacs (github.com)

Virtual Linux Devices on ARM64 (underjord.io)

Encryption made for police and military radios may be easily cracked (wired.com)

Lurk – A Turing-complete programming language for ZK-SNARKs (github.com)

Gold futures jump to record high after US tariffs on cast bars – Gold (theguardian.com)

uses this (usesthis.com)

HTTP Is Not Simple (daniel.haxx.se)

OpenAI bringing back GPT-4o to ChatGPT Plus users (old.reddit.com)

Jim Lovell Has Died (en.wikipedia.org)

Turn any website into an API (parse.bot)

What Is Popover=Hint? (una.im)

Achieving 10,000x training data reduction with high-fidelity labels (research.google)

Foundry (YC F24) is hiring staff-level product engineers (ycombinator.com)

A love letter to my future employer (2020) (catzkorn.dev)

Cursed Knowledge (immich.app)

Writing a storage engine for Postgres: An in-memory table access method (2023) (notes.eatonphil.com)

Ask HN: Is it even worth selling to Europeans given the new AI regulations?

Comments (11)