Agents with tools
Agents that plan multi-step work, act through defined tools and APIs in your systems, and report back, with every step logged.
Services / 03
AI & Custom Models
Custom AI for organisations with specialist language, data or workflows: fine-tuned models, evaluation, private deployment and the product around them, built on our own NVIDIA hardware in Athens.
What it is
We train models. Then we ship them.
A custom model is an AI model adapted to a specific domain, language or task, usually by fine-tuning an open-weight model on data you have the right to use. It learns your terminology and formats, can run privately, and is measured against a baseline before anyone relies on it. We build custom models when general-purpose ones fall short, and integrate frontier models when they are the right tool. Med-1, the Greek medical speech model at the core of Omnimedica, was developed by Designco.
What we build.
Intelligence with a defined job, a measured baseline and a product around it.
-
Domain-specific language models.
Models fine-tuned on your terminology, documents and formats, for tasks such as drafting, extraction, classification and search.
-
Speech and dictation.
Speech models for specialist language, like Med-1, which handles Greek medical dictation inside Omnimedica’s reporting platform.
-
AI inside your products.
Frontier models integrated into platforms and internal tools, with permissions, logging and human review designed in.
-
Creative AI workflows.
Image, video, audio and 3D generation connected in FlowNode, our own platform, with brand kits and custom model training.
-
Evaluation and deployment.
Task-specific evaluations, regression tests, quantisation and private deployment, so a model’s performance is measured before and after launch.
Intelligence, built in
Agentic infrastructure, built to be trusted.
Beyond the model, we build what agents need to work safely inside a business: tools, permissions, memory, evaluation and a person in the loop wherever the stakes call for one.
Conversational interfaces
Chat and voice designed as carefully as any screen, grounded in your data and able to hand over to a person at any point.
Guardrails and review
Permissions, approvals, evaluation and monitoring designed in, so an agent’s autonomy grows only as fast as the evidence allows.
Models in use.
From the model to the product around it.
How we work
Four stages.
-
Define the task
We agree what the model must do, how success will be measured and which data can legally and safely be used.
-
Prepare the data
Curation, cleaning, labelling and versioning, with sensitive information handled deliberately and leakage kept out of the test sets.
-
Train and evaluate
Supervised fine-tuning and parameter-efficient methods such as LoRA and QLoRA, run on our NVIDIA DGX Spark systems and compared against strong baselines.
-
Deploy and monitor
Inference, latency and cost tuned for production, a documented handover and ongoing evaluation as the data and the use change.
Scope
What you get.
Deliverables
- AI strategy & feasibility
- Dataset curation
- LLM fine-tuning
- Speech models
- Evaluation suites
- AI integration
- Conversational interfaces
- Private deployment
- Model monitoring
- Documented handover
- Agentic infrastructure
Tools & hardware
- NVIDIA DGX Spark
- PyTorch
- Hugging Face
- NVIDIA NeMo
- LoRA & QLoRA
Sectors
- Healthcare
- Finance
- Shipping & maritime
- Public sector
- Creative teams
Questions
Asked and answered.
Do you build AI from scratch or use existing models?
We choose deliberately. Most custom models start from a strong open-weight model adapted to your domain, and many products are best served by a frontier model through an API. Either way, we do not claim an improvement we have not measured against a baseline.
What is agentic infrastructure?
The tools, permissions, memory, evaluation and oversight that let AI agents carry out multi-step work inside your systems safely. We design and build it together with the agents themselves.
Where does our data go?
Only where we agree it can. We work with data you have the rights to use, train on our own hardware in Athens when the data should stay under our control, and deploy privately when your requirements call for it.
Can AI make decisions in our workflow?
It can prepare, suggest and automate, but we design for accountable people. In Omnimedica, for example, clinical interpretation and final approval stay with the professionals using it.
How do you know a model works?
We build task-specific evaluations and held-out tests before training, compare results against baselines, and keep regression checks running after launch.
What hardware do you use?
We fine-tune and evaluate on in-house NVIDIA DGX Spark systems, and deploy to the infrastructure that suits your requirements.
More services
The rest of the house.
- 01
Enterprise Platforms & Portals
Customer portals, partner platforms, multi-tenant SaaS and the internal systems behind them.
- 02
CRM & ERP Systems
Custom CRM, ERP and operations software: customers, orders, assets, fleets and reporting in one system.
- 04
Transactional Systems
Leasing platforms, booking engines, quoting, applications, payments and commerce.
- 05
Product & Experience Design
UX research, information architecture, UI design, prototyping and design systems for complex software.
- 06
Brand & Identity
Positioning, naming, visual identity and brand systems that carry into products and websites.
- 07
Immersive Web & 3D
WebGL and Three.js websites, real-time 3D, product visualisation and 3D production.
- 08
Security & Governance
DevSecOps, manual code audits, row-level security, multi-tenant governance and compliance by design.
