NoemaAI for organizations
Custom model pilots
A guided engagement to adapt a model with data you approve, for your terminology, output formats, and workflows, then evaluate it on your task and your hardware.
What a pilot can target
One workflow at a time.
Pilots start from an appropriate pretrained model, including a Noema model when it fits. We record the base model, the method, and the resulting artifact at every step.
Classify company-specific requests or documents
Extract into your required fields and formats
Draft in approved terminology, structure, and style
Use permitted tools more reliably in one workflow
Example workflows
Where private AI earns its place.
Illustrations of work we can scope with you, not published customer results. Each one starts with a named task and a measure agreed before any change is made.
Support request triage
Classify incoming internal requests and produce a structured handover for the right team on a company Mac.
- Approach
- Adaptation on approved historical examples
- Measured by
- Correction time and label accuracy
Onboarding and reference
Answer new employees’ questions from the handbook and current procedures, with the sources shown.
- Approach
- Retrieval with role permissions
- Measured by
- Answer correctness and citation quality
Document extraction
Pull the fields your systems need from forms, reports, or contracts into your schema.
- Approach
- Configured prompts, then fine-tuning if needed
- Measured by
- Field validity and format adherence
Drafting in house style
Draft replies or reports in approved terminology and structure, grounded in current material.
- Approach
- Retrieval combined with adaptation
- Measured by
- Human edit time and terminology use
Retrieval or adaptation
The method follows the task.
Retrieval supplies current facts but can miss the right source. Fine-tuning can improve behavior but can also introduce regressions. We choose from an observed need and the data available, then evaluate the complete workflow.
| Need | Initial approach | Evidence before expanding |
|---|---|---|
| Answer using changing internal information | Retrieval with permissions and source references | Answer correctness, citation quality, and access tests |
| Follow a consistent style, schema, or procedure | Configured prompts; fine-tuning on approved examples if needed | Improvement on unseen cases against the configured baseline |
| Improve a specialized task or domain vocabulary | Task fine-tuning or domain adaptation | Representative data, a defined task metric, and retained general behavior |
| Reduce latency or memory for a specific workflow | Model selection, quantization, and possibly distillation | Quality and performance measured on the target device |
| Use current knowledge with specialized behavior | Retrieval combined with model adaptation | Evaluation of each component and the combined result |
How a pilot runs
Measured before it ships.
- 01
Discovery
Identify the task, the current process, the available data, and your hardware.
A specific problem suitable for evaluation - 02
Scope
Agree deliverables, acceptance criteria, processing location, timeline, and budget.
An agreed pilot scope - 03
Baseline
Evaluate the configured existing model and workflow before changing anything.
A measured reference and its limitations - 04
Adaptation
Prepare the approved data and perform only the customization the baseline justifies.
A reproducible candidate and run record - 05
Validation
Test held-out cases and measure size, memory, and latency on your target devices.
A comparison against the agreed criteria - 06
Deployment
Package the approved model and deliver it to the permitted pilot audience.
Use by the pilot group - 07
Support
Handle updates, regressions, and compatibility for the deployed model.
A defined maintenance plan
If adaptation does not improve the agreed outcome, we say so. You receive the measurements and a recommendation for the configured baseline or another approach. An evaluation can end in a clear decision without a new model.
Deployment and data handling
Local use and training are separate.
In a local deployment, inference runs on your supported devices. If you commission model adaptation, we agree which data may be processed, where training takes place, who can access it, and how long it is retained. Client data is not used for unrelated models or other customers without explicit permission.
Normal local use
Inference on your devices
With Noema 4.2 or later, approved devices can retrieve and generate locally. Direct retrieval contacts your company connector; remote inference requires a company-approved endpoint. Protected content is not processed by Noema servers.
Optional adaptation
Approved data, agreed environment
Only data you explicitly approve is processed, in the training environment agreed for the engagement. Access, retention, and deletion of files, checkpoints, and logs are set in advance.
Administration
What the service needs to operate
Workspace accounts, billing, console roles, device-delivery registrations, audit records and encrypted artifacts. Noema does not hold the customer signing or decryption keys. Optional model adaptation has a separate agreed data scope.
Permissions apply before training. A model can retain what it was trained on, and role-filtered retrieval cannot remove it. A broadly distributed model is trained only on data approved for that audience; restricted information stays in a separate model for its permitted audience or in controlled retrieval.
Commercial approach
Scoped projects, ongoing support.
We propose pricing once the scope is clear.
Evaluation and adaptation
A scoped fee for evaluation, data preparation, and adaptation, with included iterations, deliverables, and acceptance criteria written down.
Deployment and support
Recurring Noema Teams deployment, administration, and support where your organization needs it.
Further work
Retraining, integrations, and extended support are scoped separately when you need them.
Our work
What the starting point looks like.
Customer results appear only from real engagements, with permission. Until then, this is our own work.
Open-weight models
Noema 1.5 4B, Noema 1.5 2B
Post-trained from Qwen3.5 bases and evaluated against them, with methodology on each model page.
Browse the modelsResearch · Preprint
Reusing Model Probabilities for New Queries
Related work on reusing model probabilities across label groupings, including intent classification.
Read the paper
Questions
Pilot questions.
Is custom model adaptation available now?
Yes, as a guided pilot service. Each pilot has a defined scope, deliverables, and acceptance criteria, and NoemaAI does the work hands-on with you. There is no self-service training dashboard.
What happens to data we provide for adaptation?
We agree in writing which data may be processed, where training takes place, who can access it, and how long source files, checkpoints, and logs are kept before deletion.
Who owns the adapted model?
Usage rights, export, and retention for each artifact are agreed in the engagement’s scope. The base model’s license still applies, and we record the base model and method for every artifact we deliver.
Can each employee get a personalized model?
Pilots start at the organization and task level: your vocabulary, templates, procedures, and approved examples. Role-specific prompts, knowledge access, and model assignments add further personalization. Per-employee training and continuous learning from conversations are not offered.
How much does it cost?
Pricing depends on the task, data, devices, and support you need. We propose it once the scope is clear, with the deliverables and responsibilities written down.
Start a conversation
Discuss a pilot.
Email us a short description and we’ll follow up with a practical next step. Please do not send confidential datasets; data transfer follows an agreed scope and handling process.
Helpful to include
- Your organization and a business contact
- The workflow you want to improve and who would use it
- Your devices: iPhone, iPad, Mac, and their mix
- Deployment or security constraints
- What interests you: Teams, company knowledge, a custom model, or help choosing
- A description of the data available and any processing restrictions