A credible baseline
Use your best prompt and current model configuration, not a deliberately weak starting point.
Custom Model Sprint
A custom model trained on your conversations.
We prepare your training data, fine-tune a model, and test whether it gives the replies you want. You get the dataset, an adapter you can run, and the test results. Projects to tune conversational style typically cost $25,000-$80,000 and take 4-6 weeks.
Why hire us if your team can run a LoRA job?
Running the job is only part of the work. We help your team choose the examples, agree what good replies look like, and check whether training improves them without breaking other tasks.
Data worth training on
Select representative conversations, approve the target replies, and keep evaluation conversations separate.
A release decision
Score target behavior alongside safety, task completion, and capability regressions before recommending rollout.
Files your team can run
Receive the agreed model files, test results, and setup instructions.
What the sprint covers
We focus on one conversation task and one model setup. Before work starts, the statement of work sets out the data, experiment budget, deliverables, and checks used to accept the work.
Review the problem
Look at failed conversations, agree how to score replies, and test whether a prompt change, better examples, or a tool fix is enough.
Prepare data and train
Have reviewers approve example replies. Train an adapter if the initial review shows it's worth testing.
Test and hand over
Compare replies on conversations kept out of training. Check for new failures in your setup and recommend whether to deploy, revise, or stop.
A typical 4-6 week schedule
The schedule depends on how ready your data is, how quickly reviewers can respond, and the integration work. We agree project dates after the initial review.
Weeks 1-2: Review
Agree data access, choose the replies to improve, test the best prompt, and set aside test conversations.
Weeks 2-4: Prepare and train
Review example replies and run the agreed training experiments. Keep changing facts in documents or tools the model can look up.
Weeks 4-6: Test
Compare results, document new failures and untested cases, and check the settings your team will use.
What your team receives
The statement of work lists the file formats and acceptance checks. A training project includes:
Review and test criteria
Recurring mistakes, scoring rules, the existing setup, and our reasons for recommending training or a simpler fix.
Reviewed dataset
Approved reply examples linked to their sources, plus separate test conversations and a record of which examples went into each set.
Custom adapter
When training goes ahead, the adapter weights for your chosen base model, training settings, and selected checkpoint.
Test report
Side-by-side replies, scores, and any drop in other capabilities. Response-time and cost measurements are included where agreed.
Run instructions
The software, prompt, chat template, quantization, and adapter settings used in testing, with instructions to reproduce the setup.
Deployment recommendation
Which tests passed, what still fails, and whether to deploy, revise, or stop. Better test scores don't guarantee better business results.
Free template
Get the trace readiness checklist
Confirm you have the export fields and reviewer access needed before a sprint kickoff.
This template uses made-up examples, not client data. Adapt it to your policies, privacy requirements, and approval process.
Who this is for
- Founders, CTOs, and ML leads building character, companion, or voice products
- Teams with permission to use their conversation data and a recurring problem to fix
- Teams with a reviewer who can approve the replies used for training
- Support or other chat systems with a specific task we can test
Who this is not for
- Teams without production traces or reviewer access
- Buyers seeking guaranteed outcomes before sharing data
- New demos with no real conversations to review
What determines the price
Dataset work
Volume, cleanup, annotation depth, and how much reviewer time the target behavior requires.
Training scope
Base-model size, experiment count, compute budget, and capability checks. Included compute is specified in the quote.
Deployment and integration
The software and quantization you need, voice-system tests, and handover work. Hosting and ongoing operations are priced separately.
What we need from you
A sample of conversations you have permission to use, your current prompt and model setup, examples of good replies, and a reviewer who can approve corrections. Remove unnecessary personal details. Before you send data, we agree how it will be transferred, who can access it, how it may be used, and when it will be deleted. Plain conversation exports are enough.
How we test changes
We review real conversations with your team and agree what the agent should do differently. Approved corrections can become training data, test cases, or examples the agent looks up. We compare each change with your best existing setup on separate test cases, then check for new failures in the software you plan to deploy.
References:
Common questions
How do we agree on success?
Before training, we agree the test cases, scoring rules, failures that must not happen, and any acceptable loss in other capabilities. We also set response-time or cost limits where needed. We compare with the best prompt we agree on. To test preference, reviewers judge replies without knowing which model wrote them. Shorter replies alone aren't success.
What does the 2x2 comparison mean?
We run four tests on the same messages: the base and tuned model, each with and without the prompt. This helps show what the prompt and adapter each change. The final deployment decision depends on separate test conversations using the prompt and software you plan to deploy.
Who owns the data and adapter?
The contract sets out ownership, usage rights, and delivery formats for the dataset and adapter. The base model keeps its own license. We check its restrictions with you before choosing a model; delivering an adapter doesn't remove them.
How is conversation data handled?
Remove unnecessary personal data or mask it before transfer, and confirm you have permission for the agreed use. Before work starts, we agree in writing who can access the data, where it is hosted, how long we keep it, when we delete it, and whether any reuse or publication is allowed.
What if training is not the right fix?
We'll recommend a prompt change, retrieved examples, or a tool fix if that is enough. The statement of work says when we make that decision and how the work and fees change if training stops. We don't guarantee improvement.
Do you host the production model?
We hand over the agreed adapter and settings for your infrastructure or chosen host. Hosting, monitoring, maintenance, and extra integrations are separate work and must be agreed separately.
Related
Discuss your model and the replies you want to improve.
Bring a few conversations with personal details removed and your current prompt. We'll discuss the data, tests, and work needed for one task.