labs@atheron:~/services/ai$ agent run --tools --eval
AI agents and models
Real work, within limits you set.
We build on every frontier model, on open-source models run wherever you choose, or on a custom LLM we create for you. We help you pick the one that fits your data and your budget. We train our own models on our own GPUs, so our advice comes from doing it.
// what we build
One assistant, or many agents.
01
Any frontier model
02
Open-source models
03
Agentic systems
04
Retrieval (RAG)
05
Fine-tuning
06
Custom LLMs
07
Inference
08
Evaluation
// how we build it
Limits in the tools themselves.
- We build with agents too. AI agents, directed and reviewed by our engineers, lay the foundations of every project, which keeps our senior engineers on the critical work.
- An agent can only do what its tools allow. We write your limits into the tools, so a clever prompt cannot talk its way past them.
- Every agent gets an evaluation set before it gets users. Every change is measured against it.
- Personal data stays out of logs. Conversations are kept only when there is a reason to keep them.
- Costs are counted per call, with a daily cap, so a busy day never turns into a surprise invoice.
// stack
What we usually build with.
- models
- Every frontier model, open-source models on your hardware, or a custom LLM trained for you
- training
- PyTorch, XGBoost, CUDA, Hugging Face transformers
- retrieval
- PostgreSQL with vector search, or a dedicated index where scale needs one
- agents
- Tool use with typed inputs, streaming, prompt caching and spend caps
- hardware
- Cloud GPUs, or Lenovo ThinkSystem GPU servers we specify and set up
// a typical first project
An agent pilot on your own data.
One agent, three or four of your tools, retrieval over your documents, and an evaluation set. Open it in the estimator to see what it would cost.
// proof
ATON AI and our checks
A trading model we train on our own GPUs, a news labeller that runs locally, and the automated checks that review our code every four hours.
// questions
Questions about AI projects.
Will our data be used to train someone else's model?
Not by us. We choose providers and settings that do not train on your data, or we run open-weight models on hardware you control.
Do we need our own GPUs?
Usually not to start. Hosted models and rented GPUs cover most pilots. If you train often or must keep data on site, we can specify and set up a GPU server for you.
How do you know the agent is right?
We build an evaluation set from real examples first, then measure every version against it. You finish a pilot with numbers, not a feeling.
What does it cost to run?
It depends on the model and your traffic. We estimate it with you, cap it, and use cheaper models for the steps that do not need the best one.
// industries
How it fits your industry.
See what this looks like in your industry: the questions buyers ask, and example projects with a price range.
// related
Often built together.
// where we work
Find your market.
See this work from your city: what changes there, the questions buyers ask, and prices in your currency.
// start
Not sure which you need?
Tell us the problem, not the solution. We will tell you what we would build, and what we would not.