Narufield AI

Where to Start

What I Build

How it goes

Salience Engineered

It said something it shouldn't.. it's inconsistent.. it forgot.. we can't make it sound like us.. it's plain out wrong... sound familiar?

Getting AI to behave coherently over time is hard for everyone who tries it, “AI Experts” included. You are trying to shape the output of a statistical machine, and usually you do that by prompting it: what to do, what stance to take, how to play a persona. Those requests are left to the model's interpretation, and they get messy fast as complexity rises.

My approach to shaping behaviour targets a more first-principle lever, what a model attends to and how. Against traditional prompting this discipline gives me surgical control across model families, with smaller instruction sizes and no training requirements falling out as a byproduct.

Which at the end of the day means smaller compute bills and faster changes.

Talk to some written with it here: Open Chat →

  1. Everything I build can be deployed to the EU cloud, or run locally on your own hardware. Narufield itself sits on Cloudflare in the EU; a build for you can go to StackIT in Heilbronn, to IONOS, to the Open Telekom Cloud, to Hetzner, or onto a machine in your own server room.
  2. None of it needs your data to train on. What passes through stays transient and anything stored is encrypted at rest.
  3. The systems get secured all the way through: least privilege between the parts, nothing trusted for sitting on the inside, and prompt injection in the threat model from the start.
  4. Set up the way your compliance people want to find it: encrypted at rest and in transit, GDPR handled properly, and controls aligned with ISO 27001.
  5. Systems are built to stay out of the AI Act's high-risk category. If what you want would land in Annex III, for instance because it screens applications, you hear that before anything is built.

Write to me

English or German, whichever you prefer.

Who you’d be working with

I’m Joseph, currently based in Mössingen, and I’ve been building real-time systems for 13+ years and working with language models since 2023. I’m an autodidact, self-taught, which in this line of work is how most people learn these things.

I design and develop the whole picture, from the server it runs on to the way it speaks to your staff. At this point in my career I’m looking to apply my trade in a useful manner, and right now the Mittelstand is underserved by what modern technology can do.

Better with two

Everything on this menu can run anywhere in the country. Getting it there takes someone who can talk to people the way I can build systems. If that’s you, or you know someone who is: Read on →