Artificial Intelligence
We run open-weight language models on your own infrastructure, fine-tune them on your data, and wire them into your product. Your data stays inside and the cost stops being a moving target.
If any of this sounds familiar
- We cannot send our data to a third-party API, and we do not know what the alternative looks like.
- Token costs have become unpredictable — the bill scales with usage in ways we cannot forecast.
- We picked a model but could not work out how to run it in production.
- A general-purpose model does not perform well enough on our domain vocabulary.
- We cannot decide where in the product AI actually belongs.
What's included
Self-hosted model serving
Running Gemma-family and other open-weight models on your own hardware. Your data stays inside, and cost becomes a fixed line rather than a usage curve.
Model selection & sizing
Choosing the right model family and parameter size for the job. Picking a larger model than the task needs is the most common and most expensive mistake.
Fine-tuning
Adapting a model to your own data: dataset preparation, training, evaluation, and comparison against the previous version so you know whether it actually improved.
Product integration
Wiring the model into your product — API layer, prompt design, context management, and sane fallback behavior when the model is unavailable or wrong.
Reinforcement learning in Unity
Training agents in game and simulation environments: environment design, reward shaping, the training loop, and embedding the trained policy back into the build.
GPU capacity planning
Working out the hardware required for your expected concurrency, and the honest arithmetic between owning GPUs and renting them.
Monitoring & cost tracking
Latency, throughput, and resource utilization under observation. A model server is a production system and needs the same operational discipline as one.
Handover & documentation
Model versioning, retraining procedures, and runbooks delivered in writing.
Technologies we use
- Gemma
- Fine-tuning
- RL / ML-Agents
- Unity
How we work
Feasibility & prototype
A working prototype that answers whether the idea holds up — before a large commitment.
Build project
Standing up model serving on your infrastructure, integrating it into the product, and taking it to production.
Ongoing retainer
Model updates, retraining cycles, and performance tracking. This field moves quickly; we stay alongside it.
Frequently asked questions
What hardware investment does self-hosting require?
It depends on model size and concurrency. A mid-size model serving a bounded number of concurrent requests fits comfortably on a single GPU. We calculate the real requirement against your usage pattern on the first call — before you buy anything.
Are open-weight models as good as commercial APIs?
For open-ended general chat, the frontier commercial models are still ahead. For a narrow, well-defined task — document classification, question answering over your own corpus, structured extraction — an open model fine-tuned on your data is usually sufficient and sometimes better. And your data never leaves your infrastructure.
How much data does fine-tuning need?
Less than people expect. A few thousand well-chosen, consistently labelled examples make a meaningful difference for most narrow tasks. Data quality matters far more than volume, which is why we start by reviewing the dataset rather than the model.
We want to add AI to our product but do not know where to start.
This is the most common question we get. We usually begin with a short discovery: separating the places where there is real measurable gain from the places where there is only demo value. If AI is not the right tool for something, we say so.
What is reinforcement learning in Unity actually for?
It is used where hand-written rules run out: opponent and companion behavior learned through training rather than scripted, or policy optimization inside a simulation environment. It is the point where our game work and our AI work meet.
Related services
Tell us what you're trying to build.
The first call is technical — no slide deck. We listen to what you're running and tell you whether it's workable. We reply within 2 business days.