Proven with eval receipts — no marketing math

Your agent, trained on your domain.

Tell us what your business does. We build the model that makes your assistant reliable — then you plug it in, or we do it for you.

What we can do for you

1 live · 3 built to order
Live

Your support agent should never do the wrong thing

When a customer asks to freeze a card or move money, the assistant has to pick exactly the right action — and never one the customer isn't allowed to take. Ours gets it right 98.2% of the time and has never once exceeded a permission level in testing.

For developers: model ids bladelabs/tool-router-qwen3-4b (hosted) and bladelabs/functiongemma-270m-router (edge, 270M).

Built to order

Insurance claims that route themselves — correctly

Your intake assistant should recognise the claim type, gather the right facts, and hand it to the right team on the first try. We build a model tuned on your claim shapes, prove it against a held-out test set, and hand you the score before you pay.

Built to order

Telecom support that goes to the right desk

Fewer transfers, faster resolutions. We train a model on your ticket history and tool set so the assistant knows which system to touch for each request — and can prove it on your own data.

Built to order

Something else entirely?

Tell us the job. We build a validated, domain-tuned assistant in days — with a scoreboard that shows how it does on real cases, before you commit to using it.

Advisor

Not sure which API fits?

Describe what your agent needs to do. We'll tell you whether the live Banking Tool Router covers it, or whether it's a built-to-order job.

0 / 500

AI-generated starter — verify against the docs.

The receipts

Every claim carries a scoreboard.

Held-out test of 997 real banking requests. Our tuned model picks the right action 41.6 points more often than a 120B general-purpose model — and has never once acted above what the customer was allowed to do.

view full scoreboard
right action chosen997-case test
off-the-shelf small model47.2%
120B general model56.6%
Blade tuned98.2%
0 permission violations across the test

How it works

01
You describe the job

Tell us what your assistant needs to do, what it's allowed to do, and where it plugs in.

02
We build and prove the model

We tune a model for your domain and hand you a scoreboard on realistic cases — before you use it.

03
You use it — in your product, or through ours

Plug it into what you already have, or let us run the assistant on your behalf.