Custom models for voice AI agents.

Trained on your calls. Held to your rules.

We build the dataset, train the model, and run it for you.

Trusted by top Voice AI engineering teams

Time to first token: 56 ms
Shorter prompt: ~8×
Critical rules broken: 0

A prompt can't hold every rule. A model trained on them can.

Your agent's rules live in a prompt thousands of tokens long, sent on every turn. The model re-reads them, still slips on the rare cases, and makes every caller wait while it does. We train the rules into a small model instead, so it holds them, answers sooner, and can be yours to keep.

It holds every rule
A prompted model re-reads your rules on every turn and still slips on the rare cases. Even GPT-5.4 broke critical rules with the same prompt. Trained in, they hold.
It answers without the wait
Prompt size is one of the biggest drivers of response time, and you pay to resend it on every reply. Without it, replies start sooner and the same instructions stop costing you.
It’s yours to keep
A rented model can change under you. A trained model can be yours, and run wherever your compliance needs it.

Step 01 · Benchmark

Your rules become a benchmark.

Every version we train is scored on it, and only the best one ships.

Your input
Your team tells us the rules and edge cases that matter. Each one becomes a family of cases.
Production calls
When a new failure shows up on your live calls, we cut that moment into a new case.
Red teaming
Simulated callers try to break your rules on purpose. Whatever gets through becomes a case.

BaseBase
v1Blocked
v2Blocked
v3Blocked
v4Blocked
v5Ships

Step 02 · Dataset

Your calls become the dataset.

Real traffic is lopsided. Before training, we even out your calls so the model sees every situation enough times.

01
Your traffic, as it comes
A few situations fill most of your calls. The hard ones barely show up.
02
Trim and repair
We sample down the common cases so they don't drown out the rest, and repair calls that broke a rule.
03
Fill the gaps
We write calls for the rare cases, aimed at the rules your model gets wrong on the benchmark.

Real anonymized callsRepaired callsWritten calls

Step 03 · Training

We train it until it passes.

We fine-tune a model for your agent and score every version on your benchmark. It ships when it holds every critical rule.

Fine-tuned on your dataset
A small model, trained on the calls from step 02.
A note before each decision
It writes itself one line on which rule applies, then answers. Callers never hear it.
Ships only when it passes
Every version is scored on your benchmark. One that breaks a critical rule stays out.

CallerHello?
NoteAnswered with “Hello?” → introduce myself, ask who I’m speaking with. Reveal nothing about the account.
AgentHi, this is Emma from Bridgewater Finance. Am I speaking with James Carter?
CallerCan I pay it on Friday the 30th?
NoteProposes Fri 30; deadline is Wed 28 → don’t accept, offer Wed 28.
AgentI’m sorry, the latest I can do is Wednesday the 28th. Would that work for you?

Step 04 · Inference