Custom AI

Your data. Your workflow.
Your model.

When the six agents don't fit, we build one that does: a custom agent on your systems, or an open-source model trained on your data and run on your servers.

See how we build it

code and weights you own: 100%; code and weights you own
license or per-token fees: $0; license or per-token fees
for training and running: Your servers; for training and running

Two ways to build

Start with the problem. We'll pick the build.

For a workflow only you have

Custom AI agent

A new agent built on the same connected layer as the six. It reads your ERP, PIM and CRM, does the work, and waits for approval.

Examples

RFQ to quote: a 140-line RFQ drafted in 15 minutes instead of 3 hours, with 9 lines flagged for a rep.
Rebate recovery: $212,000 of unclaimed supplier rebates found across 9 programs.
Returns triage: 6 in 10 returns resolved in one touch.

Also: credit holds, supplier onboarding, warranty claims.

Typical build4–8 weeks

For a task generic AI gets wrong

Custom model

An open-source model trained on your own records, so it understands your part numbers, abbreviations and rules. It runs on your servers and you keep the weights.

Examples

Part number cross-reference: 94% of competitor and OEM part numbers matched on the first try, up from 60%.
Technical documents: "Which valves are rated for 300 psi steam?" answered with the spec sheet and page.
Inside sales inbox: 1,800 emails a week sorted, with PO lines pulled into draft orders.

Also: messy descriptions to clean records, forecasting for your SKUs.

Typical build6–12 weeks

Figures are illustrative, based on a sample distributor.

Why train your own

Generic AI reads the words. Your model knows what they mean.

InputSS 1.5HP CENT PMP 230/460 NSF

Generic model

A stainless steel pump, 1.5 horsepower, possibly centrifugal. Voltage and certification unclear.

71%fields correct

Model trained on your catalog

Product class: Centrifugal pump
Housing: 316 stainless
Motor: 1.5 HP
Voltage: 230/460 V
Certification: NSF/ANSI 61
Matches: PMP-441

97%fields correct

Illustrative example on sample catalog data.

What you get

Small, specific and yours.

Speaks your language

Part numbers, supplier codes, house abbreviations. Learned from your records, not guessed from the internet.

Stays inside your walls

Training and running both happen on your servers. Your data is never used for anyone else.

No per-token bill

A smaller open-source model on your own hardware. Usage goes up; the bill doesn't.

Gets better with use

Every correction your reviewers make becomes training data for the next version.

How we build it

One workflow. Proven before it scales.

01
Discover
Pick one workflow. Agree what good looks like and measure how it runs today.
02
Prepare
Pull and clean the data from your systems. You approve exactly what is used.
03
Build
Build the agent or train the model, then test it on your records it has never seen.
04
Prove
Run it side by side with your current process. Go or no-go on the number we agreed.
05
Hand over
Deploy on your servers. Code, weights, prompts and tests transfer to you.

What you own

Everything we build.

Source code and workflows
Model weights
Data pipeline and training scripts
Evaluation set and test results
Prompts and documentation
No license fee, ever

How it behaves

Built to be trusted.

Nothing writes back without approval
Every answer traces to its source records
Tested on your data before go-live
Full audit trail of decisions
Says 'I don't know' instead of guessing

Before you ask.

How accurate is it?

Accuracy on a test set isn't enough. It has to hold up on your live systems. We work with you on your own data and prove it on your use case, running against your production systems.

Which models do you use?

Open-weight models licensed for commercial use, such as Llama, Mistral or Qwen. We pick the smallest one that hits your accuracy target, because smaller is faster and cheaper to run.

How much data do we need?

For a narrow task, often a few thousand good examples. Discovery tells you whether you have enough before you commit to the build.

Where does it run?

In your cloud account or on your own hardware. Training happens there too. Nothing is sent to us or to a model provider.

Agent or model: how do we choose?

If the work is a process across your systems, it's an agent. If generic AI keeps getting a specific task wrong, it's a model. Many projects use both.

What does it cost?

A fixed quote after discovery, before any build starts. After go-live you pay only for your own servers.

What if it isn't accurate enough?

Then we stop at the Prove step. You only continue if it beats the number we agreed at the start.

Bring one workflow you'd fix first.

Thirty minutes. We'll tell you whether it needs an agent, a model, or neither.

Fixed quote after discoveryRuns on your serversYou own the weights

Your data. Your workflow.Your model.