AI & ML interests
None defined yet.
Recent Activity
Ionate Technologies
An AI lab with an engineering studio attached.
We build custom reinforcement learning systems and autonomous AI agents, and the engineering to run them in production: web, mobile, cloud and data.
Most AI is trained once and frozen. We build the other kind: systems that learn from their environment, adapt when the world shifts, and ship as real software a team can run.
🌐 ionatetechnologies.com · Solutions · Start a project
What we work on
- Custom reinforcement learning. We design the environment, shape the reward and train policies that make decisions in your domain. Then we prove they hold up outside the simulator.
- AI agents & automation. Agents that plan, call tools and coordinate with each other, built with evaluations and human checkpoints so you can trust what they do.
- Data & ML infrastructure. Pipelines that arrive on time, versioned datasets, and training and serving infrastructure that's cheap to run.
- Engineering around the model. Products, on-device inference, cloud, embedded systems and integrations: the rest of a working system.
Live in your browser
Each of these trains or solves in real time on our site, with nothing pre-recorded. Simulated inputs (rooms, shoppers, services) are labelled as simulated, but the algorithms are real.
| Demo | What runs |
|---|---|
| Watch an agent learn | Tabular Q-learning vs SARSA on a grid you can redraw (α 0.5, γ 0.95, ε 1 → 0.05) |
| Watch a network learn | An MLP with two hidden layers of 12 tanh units, cross-entropy loss and Adam, plus a live decision map |
| Teach a phone a gesture | An on-device classifier trained live on $1-style stroke features; teach your own gestures, with an int8 quantisation comparison |
| Train a recommender, live | Matrix factorisation by SGD, held-out error against a popularity baseline, and ridge-regression fold-in for a new user |
| Watch a robot map a room | Simulated lidar, a log-odds occupancy grid, scan matching, frontier exploration and A* |
| Find the root cause of an alert storm | EWMA anomaly baselines with personalised PageRank on the call graph (after MicroRCA) |
| Map two APIs automatically | Field similarity and format detection, Hungarian assignment, a generated adapter and contract tests |
| Pull a palette out of any picture | k-means++ in OKLab, with design tokens checked against WCAG contrast |
How we work
- Discover: the decision to automate, the data available, and the constraints that cannot move.
- Model the environment: a simulator or digital twin, validated against reality.
- Train and iterate: reward design, algorithm selection and fast training loops.
- Evaluate: held-out scenarios, stress tests and adversarial cases. Nothing ships on a single good curve.
- Deploy and monitor: shipped behind guardrails, with drift monitoring and retraining.
Every system ships with evaluations built from real cases and explicit guardrails. It's handed over with the code, infrastructure as code and runbooks to run it without us.
Get in touch
Tell us what you're trying to automate, optimise or learn: ionatetechnologies.com/contact