Custom AI Models · LoRA · MLX · Private AI

AI model fine-tuning for your business

We build specialized models around your business data, terminology and workflows. We first assess whether fine-tuning can improve the outcome, rather than training without a clear reason.

Discuss your use case

What is AI model fine-tuning?

Fine-tuning continues training a model on examples selected for a specific task. LoRA adapts a small set of parameters through adapters rather than updating the entire model. At improveIT, we assess data and compare prompting and RAG first, then recommend training when it offers a measurable benefit for the agreed task.

Example business use cases

  • Classification and extraction — Repeatable labels or structured document data, supported by representative examples and evaluation.
  • Response style and format — Agreed terminology, tone and output schemas when instructions alone do not provide enough consistency.
  • Specialized tasks — Adaptation for a narrow workflow; current document knowledge can still be supplied through RAG.

We deliver services and training remotely for businesses in Poland and other countries, including teams in Kraków, Tarnów and Nowy Sącz. Delivery arrangements and the exercise environment are agreed before work starts.

AI implementation for business

Specialized AI with measurable outcomes

For businesses looking to improve AI output quality, explore lower inference costs or run models on their own infrastructure. Benefits are measured rather than assumed.

Prompting, RAG or fine-tuning?

Prompt engineering

When a model needs clearer instructions, examples or output formatting.

RAG

When answers need current documents, source attribution and access-controlled retrieval.

Fine-tuning / LoRA

When repeatable behavior, style or a specialized task calls for high-quality examples and measurable evaluation. Retrieval and fine-tuning can be combined.

RAG

From data to a measured deployment

  1. Discovery and baseline

    Define the task, current failure modes, expected outcome, budget and go/no-go criteria.

  2. Data and permissions

    Review model and data rights, example quality, anonymization and training/validation/test separation.

  3. Training and experiments

    Select a base model and LoRA adapter. We use MLX primarily on Apple Silicon, matching model size to memory and task requirements.

  4. Evaluation and handover

    Compare quality, regressions, latency and cost on representative examples. Hand over the adapter, configuration and an agreed evaluation report.

  5. Production and maintenance

    Match format and serving to the client infrastructure. Agree API authorization, monitoring, versioning and rollback.

Private AI with controlled data access

Training can run locally or in client infrastructure. Local execution does not replace access controls, dataset protection or licence review. Data transfer, logging and retention are agreed before work starts. Fine-tuning can memorize data, so sensitive information is minimized and disclosure risks are evaluated.

Complete on-premise environment: private knowledge and AI hub

Scope and starting prices

Fine-tuning feasibility and data audit

Compare prompting, RAG and fine-tuning; review data, risks and success metrics.

from 372.50 USD excl. VAT

Proof of Concept — LoRA / MLX

Pilot one task, train an adapter and compare it with the base model on a held-out test set.

from 1,225 USD excl. VAT

Fine-tuning on client data

An agreed model and dataset, training, evaluation, results and configuration handover.

from 2,475 USD excl. VAT

Model deployment and API integration

Application integration, access control, monitoring and an update plan in the agreed environment.

from 1,225 USD excl. VAT

Introductory LoRA / MLX workshop

Team introduction and workflow demonstration; duration and scope agreed before quotation.

from 725 USD excl. VAT

Company training: LoRA with MLX — up to 4 people

1–2 days of practical workshops: data, adapter training, evaluation and serving.

from 975 USD excl. VAT

Indicative starting prices excluding VAT, subject to market validation. Scope and price are confirmed before work starts. Infrastructure, large-scale data preparation and licences are quoted separately. Fine-tuning does not guarantee improved quality or lower inference costs. USD service prices converted at 1 USD = 4 PLN. Final scope and price confirmed in your quote.

Full pricing

AI Fine-Tuning Training — LoRA and MLX

1–2 days of practical workshops for up to four people. Move from Apple Silicon setup to training, evaluating and serving your own adapter. Duration and depth are adapted to participant experience.

Workshop agenda

  1. Set up MLX LM, select a model and review its licence
  2. Prepare examples and prevent test-data leakage
  3. Configure LoRA, train and track the experiment
  4. Compare the base model with the adapter and inspect failures
  5. Run the adapter, integrate it and plan deployment

Prerequisites and preparation

Basic Python and terminal skills. Local exercises require a compatible Apple Silicon Mac; memory needs, model access and tool versions are agreed in advance. A shared demonstration environment can be arranged. Exercises use public, synthetic or approved client data.

What the team receives

The team receives exercise materials, experiment configuration and the adapter developed during the workshop. These support a pilot, rather than guaranteeing production readiness. Follow-on fine-tuning and integration for a real use case can be scoped separately.

LoRA / MLX

FAQ

How much do AI fine-tuning services cost?

Audits start at USD 372.50, LoRA / MLX pilots at USD 1,225 and client-data fine-tuning at USD 2,475, excluding VAT. API deployment from USD 1,225 is a separate stage. Infrastructure, large-scale data preparation and licences are quoted separately.

Can the model run without a cloud service?

Yes, subject to the chosen model, licence and hardware resources. Local or on-premise deployment is matched to client requirements.

Will LoRA reduce AI costs?

LoRA reduces the number of trainable parameters compared with full fine-tuning. Lower inference cost is not automatic; we evaluate the model, prompts, hardware and workload in a pilot.

Does this MLX workshop require Apple hardware?

The MLX workshops use Apple Silicon. Production serving is scoped separately; moving an adapter to another engine requires compatibility checks.

Technical documentation

Discuss your use case

Contact

Let’s define the first step for your project.

Briefly describe your goal or problem. You do not need a complete specification. This is a no-obligation enquiry.

Describe your project