07 Data, AI & platform
ML & AI R&D prototypes
Prototypes and experiments that test whether a machine learning or reinforcement learning idea works — in simulation or on offline data, with results you can rerun.
Outcome: Evidence to build, rework or stop, from experiments you can rerun.
This is research engineering for ideas that are not yet features. The idea becomes a measurable question, and a prototype is built in Python and PyTorch, then trained and evaluated in a simulated or offline environment, isolated from the systems you run.
It fits when building is expensive and the answer is uncertain: a reinforcement learning or prediction idea, a model that worked once and needs checking, or two approaches to compare. You receive measured results and a recommendation — a proof of concept, not a production system.
Typical needs
- A machine learning idea that needs evidence before a full build
- Model results that change between runs and can’t be reproduced
- A question to test in simulation or on offline data, away from live systems
What you receive
- A runnable prototype in Python, with its training environment or dataset
- An evaluation harness that compares the model with baselines over fixed seeds
- A report of results and limits, with a recommendation: build, rework or stop
Approach
How the work runs.
- 01
Frame a measurable question
The idea is written as a question an experiment can answer: the metric, the baselines it has to beat and the result that would mean stopping.
- 02
Isolate the environment
A simulated environment or an offline dataset is set up apart from production, so runs are safe to repeat and nothing touches the systems in use.
- 03
Train with fixed seeds
The model is prototyped in PyTorch and trained next to simple baselines over several fixed seeds, with pinned dependencies and tests in CI.
- 04
Report and recommend
Results are written up with their spread across seeds and what they do not establish, followed by a recommendation: build, rework or stop.
Technologies
- Python
- PyTorch
- pytest
- GitHub Actions
The engineer who does this work is the founder, Filip Volf — you work with him directly.
Questions
About this service.
How is this different from applied AI & business automation?
Applied AI puts language models into a working process, with a person approving the output. This service answers an earlier question: whether a model or learning approach works at all, in simulation or on offline data. If it does, the prototype and its report become the starting point for a production build.
Do experiments need access to our live systems?
No. Experiments run in a simulated environment or on offline data agreed with you, apart from production. In Ariaska RL, the founder’s research, the CI suite likewise runs offline, with network access and tool execution denied.
What if the results show the idea doesn’t work?
Then that is the answer — written down with the numbers behind it, and reached before the cost of a full build. The report states what the experiment did and did not establish, and whether a different question is worth testing next.
Related services
- 01
Custom software & SaaS
Outcome: One system of record instead of scattered tools.
- 05
APIs, backend architecture & PostgreSQL
Outcome: Business rules enforced once, where the data lives.
- 06
Applied AI & business automation
Outcome: Less manual work, with a person accountable for every decision.
Next step
Let’s talk about what you’re building.
Describe what you need built or fixed, how the work runs today and the constraints. You get clarifying questions or a first outline of the approach — by email, in English or Serbian.