BLOGCategory

Custom models

Book a free process review (30 min)Send us a message

Articles grouped by business problem, industry, and system type.

Fine-tuning, RAG or training your own model?

RAG gives the model current knowledge at every question, fine-tuning teaches behaviour through examples, and a small model is enough for choosing a category. First note the mistake: outdated terms are a source problem, the wrong department a behaviour problem. In a scenario of 1,000 tickets a month, RAG with the current terms cuts errors from 15% to 7%, and fine-tuning to 3%, which together save 12 hours of corrections, or €419 a month.

custom-modelsllmprivate-aifine-tuning

Jev 1.13 for business classification and routing

Jev 1.13 returns typed decisions with probabilities, so it is worth testing for routing cases. The confidence threshold decides how many cases go through automatically and how many go to an operator. In the example, after raising the threshold, 500 of 1,000 decisions pass automatically with 99% accuracy. At Syntalith an automation costs from €3,000 net, and with 1,000 cases a month at 3 minutes each it pays back in 3.4 months.

custom-modelsllmprivate-aifine-tuning

Local Jev 1.13 alternatives: Kev and TinyJev

Jev from TypeSafe AI runs through an API, while Kev and TinyJev are separate projects by other authors, inspired by a similar way of asking questions. For a company that needs local processing, this is a direction to test: the model reads text and picks a result from a fixed list, and quality is checked on your own categories.

custom-modelsllmprivate-aifine-tuning

A model for ranking catalog search results

An AI model can set the order of catalogue search results, and first you need to establish whether the right products reach the list at all. When they do but sit far down, a model that scores how well the query fits each product helps, trained on the catalogue team's assessments.

AI modelsmodel customization

Accepting a customized model for team use

Buy fine-tuning with an acceptance agreed before the project ends: a set of cases the model has not seen, criteria for correctness, acceptable errors and the person who takes over cases that need checking. A demo of a few correct answers does not tell you whether the version that goes into the application meets these conditions.

AI modelsmodel customization

Custom or general model: compare the work your team actually does

You compare a custom model with a general one on the same cases and under the same conditions: how many corrections remain for the employee, how long they wait for an answer and who maintains the extra version. A better first draft does not always justify the cost of maintenance.

AI modelsmodel customization

Custom AI models for rare operational events

A custom model for rare operational events is judged on two numbers: how many exceptions it catches and how many ordinary cases it sends to a person for no reason. With only a few examples of the exception, a test tells you little, so the help covers what can be assessed and the rest stays with rules and experts.

AI modelsmodel customization

Training material: who approves its use?

Before you adapt a model on company materials, establish where each group of files comes from and which work it was approved for. Your own approved texts usually go into the first trial, while client and partner materials wait for a decision by the person who knows the agreements. At Syntalith an app costs from €3,000 net, and at 300 materials to assess a month of 12 minutes each it pays back in 2.4 months.

AI modelsmodel customization

Business address quality: when is validation enough, and when should you assess AI?

An AI model helps with addresses when the record is correct but a customer note says deliveries go elsewhere. Rules and an address validation service catch typos and empty fields, and the model reads free-text notes and shows the basis for its flag. At 3,000 orders a month, rules and an address service cut errors from 4% to 2% and fine-tuning to 1%, saving 22.5 hours, or €785 a month.

business addressesdata qualitymodel customization

Is your company's archive ready for fine-tuning?

Readiness for fine-tuning depends on whether the examples show the work the model should do, with the information it will receive. The size of the archive does not settle it. First check what a clear instruction with sources achieves. At 1,500 answers a month, a clear instruction with sources cuts errors from 20% to 10% and fine-tuning to 5%, which together saves 30 hours of corrections, or €1,047 a month.

fine-tuningdata preparationAI models

Should employee corrections train the next model?

Employees' corrections become material for the next model version only after review. Some fix a model error, some add a new fact that belongs in the knowledge base, and some answer one recipient's special request. The process owner decides what becomes a pattern.

AI modelsmodel customization

Fine-tuning or RAG: what will improve product answers?

When AI gives out-of-date product availability, a current source is missing, and fine-tuning will not supply one. You give the model changing data from the system at every question. In a scenario of 1,200 product answers a month, a current data source cuts errors from 10% to 3%, and fine-tuning to 2%, which together save 6.4 hours of corrections, or €223 a month.

product informationRAGfine-tuning

Fine-tuning when the base model provider changes

When you change the provider of the base model, the adaptation does not move over by itself. You carry over the set of examples, the assessment criteria and the test questions, and you first check the new model with instructions and examples in the prompt. Tuning it again makes sense when recurring errors remain on the same set.

AI modelsmodel customization

The process changed. Does the model need retraining?

A process change does not always require retraining the model. If only the team name or the recipient changed, a mapping change in the app is enough, about €70 of work. If a category changed its meaning, you need new examples and retraining, about €1,082 in the scenario. Keep changing assignments outside the model, so that a reorganisation does not need training.

AI modelsmodel customization

Assessing the business cost of custom model errors

Two models can make mistakes equally often and leave a team with very different work. A false alarm costs a few minutes of reading, and a missed problem costs half an hour of clarification after fulfilment has started. At 10,000 orders a month a more cautious threshold gives 100 more alarms, but 40 fewer misses, and saves 15 hours, which is €524.

AI modelsmodel customization

Training a model from scratch or adapting an existing one

For most companies “our own model” means running an existing model in a chosen environment or adjusting how it answers, which adaptation provides. Training from scratch makes sense when you need full control over the training data and the model's development, with a budget and a team for years.

AI modelsmodel customization

Does choosing a category require a text-generating model?

A small classifier is enough when the application assigns text to a fixed list of categories and the examples show correct decisions. A large language model has the edge with changing categories and cases that need explaining. The choice is settled by a comparison on the same cases: wrong assignments, cases passed to a person and waiting time.

AI modelsmodel customization

Does your Polish contact center need a custom model?

An AI model in a Polish contact centre recognises the purpose of a call, and with a short, ambiguous statement a clarifying question helps more than a quick label. Consider adapting the model for recurring mistakes on statements whose meaning is already clear, and compare results on separate Polish conversations the model has not seen.

contact centersPolish languageAI models

What a single sentiment score misses in product feedback

AI sorts feedback on B2B products by the topics it concerns and keeps both experiences from one comment, for example praise for the manual and a complaint about setting up the machine. A single “positive” or “negative” score will not show that. Consider your own model when a ready tool keeps losing distinctions that matter for the product.

product researchB2B customer feedbackAI models

AI for catalog data: product width or package width?

An AI model helps with catalogue attributes when suppliers describe the same features differently and a name dictionary cannot keep up. With fixed supplier templates a dictionary of equivalents is enough, and the model proposes the target field with the source label for approval. At 5,000 attributes a month, a dictionary cuts errors from 10% to 5% and fine-tuning to 2%, saving 20 hours, or €698 a month.

product dataPIMmodel customization

Private models

A model on your hardware or in an EU cloud, tested on your own tasks.

We select the hardware and model, deploy it and measure quality on your examples. You can start with a computer you already own.

Price: quoted after reviewof the task, data and hardware
First step: 30 min with an engineer, €0written summary within 2 business days

Book a free call (30 min)Prefer to write? Describe the process

Times are shown in your own time zone. We work with clients across time zones.

See private models

Custom models

A model on your hardware or in an EU cloud, tested on your own tasks.

What our OpenAI and Anthropic partnerships mean for you.