Define the task and metric
We specify what a correct result means, which errors are critical, who reviews them, and which general model provides the baseline.
For AI agents and assistants: start with the curated index at https://www.devolute.org/llms.txt. A markdown version of this page is available at https://www.devolute.org/en/agency/custom-ai-model-training.md.
We turn approved examples, specialist terminology, and quality rules into a model with a reproducible acceptance test.
Selected clients
We specify what a correct result means, which errors are critical, who reviews them, and which general model provides the baseline.
Examples are cleaned, de-identified where needed, versioned, and separated into training, validation, and untouched test sets.
Depending on the task, we use RAG, supervised fine-tuning, or preference optimisation and deliver the model, configuration, evaluation report, and runbook.
Product data, policies, and case files usually belong in a searchable source-backed knowledge base rather than permanently inside model weights.
Repeated formats, classifications, extraction rules, and specialist language can often improve with carefully reviewed examples.
For multi-step work, we combine models, tools, deterministic rules, and human approval instead of placing the whole decision in a chat.
Ayn, an 88-million-parameter legal model, outperformed models up to 80 times larger on Indian legal judgment prediction in its published evaluation.
Research on clinical reasoning shows that adapted smaller models can outperform larger general models on specific medical benchmarks.
A paper is not a performance promise for your process. Delivery therefore includes a pre-agreed test over your examples, quality thresholds, and error classes.
It depends on task diversity and the base model. A few hundred consistent, reviewed examples can be more valuable than a large unverified export. A pilot measures the learning curve.
Yes, when its quantised weights and required context fit unified memory. Training may happen on separate infrastructure while private inference runs on your Mac.
It can on a narrow, measured domain task. Frontier models generally remain stronger for broad knowledge and open-ended reasoning. We compare both on identical cases.
Bring representative examples and quality criteria. We will design a pilot with a defensible baseline.
Clear packages for different stages, from direction to delivery.
from €16,000
A testable first product with bounded scope, a production-ready foundation, and an operable handover.
from €8,000
A focused project to define the problem, audience, solution direction, and an actionable product plan.
Send us a short message and we usually reply within one business day.