Draft concept for Entropi · figures in [brackets] are placeholders
NBot by Entropi

Run your product's AI on open models — without losing quality.

NBot runs the AI in your product on open-weight models — your choice of model, provider and location — at the quality your customers expect. Proven on your own evals. Held when models change.

OpenAI-compatible · Self-hosted · Any provider or your own GPUs
Your product sends a request NBot plans · routes Open model · EU cloud Step 1 · extract Open model · any API Step 2 · classify Small model · your GPUs Step 3 · personal data Open model · reasoning Step 4 · draft answer Checked against your evals passed · logged · learned
One request, four steps — each on the model that does it best.
Trusted by teams shipping AI in their product [customer logo] [customer logo] [customer logo] [customer logo]
Your situation

Your product depends on someone else's model.

Price

Their price sets your margin.

Every new customer grows a bill you don't control.

Data

Their terms decide where your customers' data goes.

And your customers are starting to ask.

Change

Their next update can break what you built.

And you hear it from your customers first.

Open-weight models would give you that control back. But on real multi-step work they don't hold the quality on their own — and making them do it is ML engineering most teams can't staff. NBot does that engineering for you, continuously.

How NBot works

NBot does the engineering open models need — continuously.

It sits between your product and the models. For every request it:

1

Understand

Reads the task and splits it into steps where that helps.

2

Route

Sends each step to the open model and method that does it best.

3

Check

Tests every result against your quality bar.

4

Learn

Learns from every outcome, so the next decision is better.

You change one URL.Your product keeps working the way it does today.
client = OpenAI(
  base_url="https://nbot.your-company.internal/v1",
  model="nbot")
What you get

Frontier-level results. Open-weight freedom.

Quality

Quality you can prove.

Your evals, your pass mark — [X]% of frontier quality on [workflow].

Freedom

Any model. Any provider.

Cloud, sovereign cloud or your own hardware. No lock-in.

Resilience

No break when models change.

A model is retired or changes price: NBot moves the work. No rework.

Cost

Lower cost per task.

Open-model prices for frontier-level results — [–X]% per task on [workflow].

frontier baseline first task [N] tasks quality on your eval cost per task
Illustration — [+X] points after [N] tasks once the first pilot reports.
It gets better with use

Smarter with every task you run.

NBot starts from a router pretrained on [N] public tasks, then learns from your own outcomes — which model and which method fit which of your tasks. The longer it runs, the better it fits your work.

Learning happens inside your infrastructure. Your data never trains anyone else's model.

Security and control

You decide where every step runs.

NBot is self-hosted. Sensitive steps run only on the models and locations you allow — including models on your own hardware. Every step records which model ran it and where, so you can show it to your customers and auditors.

# nbot policy — example
[data.personal]
run_on  = ["own-gpus", "eu-sovereign"]

[data.general]
run_on  = ["any-open-model"]

[quality]
eval    = "./evals/refunds.jsonl"
pass_at = 0.95

Self-hosted

Runs in your infrastructure or your own cloud account.

Policies per step

Personal data stays on the models and locations you approve.

Full audit trail

Which model saw what, where and when — for every step.

Your data trains nothing else

Learning stays inside your walls.

Proof on your terms

Don't trust our benchmark. Run yours.

Point your existing evals at NBot — it speaks the OpenAI API, so it's a one-line change. Compare quality and cost side by side with your current set-up, on your own data, before you commit to anything.

[X]%
of frontier quality
[Y]%
of today's cost per task
0
rework after a model change
Horizon · trace req_7f3a · support refund● live
“Customer asks for a refund on order 4471, charged twice…”
#StepModel · locationCostEval
1understandsmall · your GPUs[€]✓ pass
2read customeropen · on-prem[€]✓ pass
3check policyopen · EU cloud[€]✓ pass
4draft replyopen · reasoning[€]✓ pass
=answer shippedyour eval [96]/100[–X]%learned
Horizon, NBot's control plane: every step, every model choice, every cost.

“[Customer quote: what changed after moving to open models with NBot — quality held, data stayed, bill went down.]”

[Name], [Role] · [Company]
Fits your stack

Drops in. Nothing to rebuild.

OpenAI-compatible API

Your existing code and SDKs keep working.

Your agent frameworks

Sits under the frameworks you already use.

Your infrastructure

Single binary, Docker or Kubernetes — on-prem or in your cloud.

Horizon control plane

Every step, every model choice, every cost in one view.

Who it's for

Built for teams that ship AI in their product.

Customer support Document processing Agent workflows
You run AI inside a product or service your customers pay for.
Your model bill grows with every customer — [€5k]+ a month.
The work repeats, and the result can be checked.
You don't have an ML team to tune models — and don't want to build one.
How to start

Start with your own evals.

Your evals are the pass mark at every step — and again whenever a model changes.

1 · Free · [2] weeks

Eval run

Your evals against NBot, side by side with your current set-up.

2 · [€X] · [4] weeks

Pilot

NBot on live traffic, with your evals as the pass mark.

3 · [€X] / month

Licence

Re-checked against your evals whenever a model changes.

FAQ

Questions teams ask first.

How is this different from a model router?

Routers pick one model per request. NBot works per step, checks results against your quality bar and learns from your own outcomes.

How is it different from an agent framework?

Frameworks help you build agents. NBot decides which model runs each step of them — and keeps that choice right as models change.

Where does my data go?

Where you allow it. NBot is self-hosted; you set which models and locations each kind of step may use.

Am I locked in to NBot?

No. It speaks the OpenAI API; switching back is the same one-line change.

Which models and providers?

Any OpenAI-compatible endpoint and local models — [list].