AI engineering · billed by the hour From $20/hr

AI that does real work,
on hardware you own.

We design, build and deploy AI that does real work on infrastructure you own: autonomous agents that take actions in your tools, private search that answers from your own documents, models tuned to your domain, and the data pipelines and guardrails underneath. Senior engineering, paired with current-gen models, billed by the hour.

$20/hr AI dev Human in the loop Runs on your infra
What we build

Five kinds of AI.
One team, one rate.

Most projects use two or three of these together. All of it runs on infrastructure you own, on your model keys or open models we host for you, and all of it is billed by the same hour.

Agentic AI

01

Autonomous agents that take real actions in your tools, triage a ticket and route it, reconcile an invoice, move a record between systems, draft a document and file it. Scoped tools, a human-approval step, and a full action log.

e.g. an agent that clears the support first-pass and only escalates the tricky ones

Private knowledge search

02

RAG search that answers questions from your own documents, contracts, tickets and wikis, with citations back to the source, so people trust it. Nothing is sent to a third party; the index lives on your servers.

e.g. “what does our contract with X say about termination?” answered in seconds, with the clause

Custom & fine-tuned models

03

Open models tuned to your domain, your terminology and your tone, then deployed on your own hardware. You get a model that speaks your business and never phones home, no per-token bill, no vendor lock-in.

e.g. a self-hosted model that drafts in your house style and knows your product catalogue

Data pipelines for AI

04

The unglamorous part that makes the rest work: pulling your data out of the systems it's trapped in, cleaning it, chunking and embedding it, and keeping it fresh. Good AI is mostly good data plumbing, done once and done right.

e.g. a nightly job that keeps the search index in step with your live document store

Evals, guardrails & safety

05

How you know it works and stays safe: test suites that measure accuracy on your real cases before anything ships, guardrails that keep it in bounds, human approval on risky calls, full logging, and a kill switch you always hold.

e.g. an eval that has to stay green on 200 of your real cases before a change goes live
Start small

Your first AI build, live in two weeks.

The right way to start with AI is one use case, in production, doing real work behind a human-approval step. Not a six-month “AI transformation.” We pick the use case with you, ship it, and let it earn trust on your actual cases before it touches anything else.

Most first builds reach production in 25 to 60 hours. It runs on your infrastructure with your model keys (or open models we host for you), keeps a full log, and comes with a kill switch from day one.

Starter
AI Pilot
$20/ hour · billed as you go
  • One use case, agent, search or a tuned model, in ~2 weeks
  • Human-approval step on anything risky
  • Full logs and a kill switch
  • Runs on your infrastructure and model keys
  • 30 days of tuning on us
A typical first build runs 25–60 hours to production.
How we build it

AI built like software. Not demos.

A slick demo is easy. AI you can trust with real customers and real money is engineering: grounding in your own data, scoped tools, guardrails, a human-approval step, logging, and evals that prove it works on your cases before it goes live. That's the part most “AI consultants” skip.

You don't pay for a strategy deck or a proof-of-concept that never ships. You pay for a working system, in production, on your infrastructure, by the hour. That's how we keep it at $20 an hour without cutting the parts that keep you safe.

Senior engineer · AI-paired · guardrails + evals · runs on your infra
A demo that dazzles, then dies. Great in the pitch, ships nothing real.
Wired into your real systems. Grounded in your data, doing real work, with a log.
Answers from the model's memory. Confident, unsourced, wrong often enough to hurt.
Grounded in your own documents. Every answer cites a source you can check.
Fully autonomous, no oversight. One hallucination from a bad email to a customer.
Human-in-the-loop where it counts. It proposes; a person approves the risky calls.
Runs on someone else's platform. Your data and keys go with it.
Runs on your infrastructure. Your keys, your data, your kill switch.
How a build ships

Four steps. Trust earned,
not assumed.

We start narrow and only widen what the system is allowed to do once it has proven itself on your real work. No big-bang rollout, no “set it loose and hope.”

Map

A short call and a look at the problem. We find where AI clearly helps, the data it needs, and the calls a human must always sign off on. You get an estimate in hours.

Build

The system, its data pipeline, guardrails and evals, scoped tight, wired into your tools, running on your infrastructure. You get a progress link and a running hour tally.

Ship

It goes live behind a human-approval step, on real work, watched closely. Everything is logged and you hold the kill switch from minute one.

Tune

30 days of tuning on us. We widen autonomy only where it earns it, and hand over the code, models, logs and docs. It's yours.

A note on safety

Fast where it's earned. A human on the calls that matter.

Powerful shouldn't mean unsupervised. We give a system exactly the tools and data it needs, no more, ground its answers in your own sources, and put a person on the decisions that carry risk, money out, messages to customers, anything hard to undo. Everything it does is logged, and it can be stopped instantly. The result is speed you can actually trust.

  • Human-in-the-loop by default.The system proposes; you choose what it can do alone.
  • Grounded in your own data.Answers cite your sources so you can check them, not the model's guess.
  • Every action logged.A full audit trail of what it did and why, with prompts on request.
  • Runs on your infrastructure.Your servers, your model keys, your data. Nothing new to sign up for.
  • Scoped access, not the whole internet.It can only touch the systems and data you explicitly allow.
  • Evals before it ships.We measure accuracy on your real cases, not a cherry-picked demo.
FAQ

Common
questions.

If yours isn't here, a 20-minute call almost always answers it.

Talk to a human →
No catch, three rules. We bill by the hour and you prepay your hours. The project is scoped in writing before we start. And we pair a senior engineer with current-gen models to do the heavy lifting, so we earn a fair rate while you pay a fair one. That's the whole trick.
Five things, alone or together: autonomous agents that take actions in your tools; private RAG search that answers from your own documents with citations; fine-tuned and self-hosted models tuned to your domain; the data pipelines that clean, chunk and embed your content; and the evals and guardrails that prove it works and keep it safe. All of it runs on infrastructure you own.
No. A wrapper sends a prompt and hopes. What we build is wired into your real systems and data, grounded in your documents, measured with evals before it ships, and logged so you can audit it. It runs on your infrastructure and, where it matters, on open models you host yourself.
Three ways. Answers are grounded in your own data with citations, not the model's memory. Anything risky waits for human approval. And before it ships we measure accuracy with evals on your real cases, not a cherry-picked demo. If something breaks in the first 30 days we fix it free.
We're model-agnostic: Claude, GPT, or open models like Llama and Qwen. For sensitive workloads we run open models on your own servers or in an air-gapped environment, so no data leaves your infrastructure. Otherwise we use your own API keys, so you keep the billing relationship and the data terms.
Your infrastructure, your repo, your model keys. Code, models and configuration live in a repo you own and deploy to your servers. If you fire us tomorrow, you keep everything: the system, its data, the logs and the docs. Same philosophy as our installs.
No. AI work pairs naturally with the tools we install (Odoo, Zammad), but you don't have to run either. If you have a problem and the systems and data it touches, we can put AI on it.

Point AI at your hardest problem.

Tell us in two paragraphs about the problem that's eating your team's time. We come back inside a day with a plan and an estimate in hours, or an honest “AI won't help here.” Either way, you get a straight answer.