MediaAtlas builds AI agents for regulated industries on language models we train on your data, and runs them where your data lives, with autonomy tiers, human approval and an activity log. We don't demo agents. Seven of our own systems run on this stack every day.
Messy input on one side, systems of record on the other, rules in between. The agent sits in the middle, on a model trained for your job, and a person stands at the gate wherever the rules say so.
Agent platforms usually start at the API. We start one layer lower, with the model, and end one layer higher, with operations.
Watchdogs restart failed parts, health checks, a scheduler for unattended runs, notifications, a stop switch.
A frozen evaluation set from your cases, tool-call accuracy on held-out steps, a regression scoreboard.
Every step written down: what the agent did, with which model version, who approved it. Hash-chained.
Autonomy tiers per action class, approval by a named person in the web app, CLI or Telegram.
Native tool calling into your systems, MCP for scheduled jobs, dynamic tool sets that load only what a task needs.
An open-weight model trained on your data for your task and your tool calls. The weights are yours. AI training
Every pattern we offer is one we already depend on. Each system below runs at MediaAtlas, and each maps onto a job in your domain.
101 also over MCP for scheduled jobs
Speech in, speech out, in Slovenian, fully local. Calendar, reminders, e-mail, unattended jobs, image, video and music generation, camera vision. Failed parts come back in about 25 seconds.
In your domain: reception and call handling, field technicians, an assistant on a device with no network.
17,413 words from one sentence · 19 model calls
Researches the market with tools, then builds a complete service website in six resumable steps: structure, copy, structured data, images, contact form. A person approves before anything is published.
In your domain: product catalogues, tender documentation, reports generated at volume.
of outgoing messages approved by a person
Drafts messages, sorts the inbox, drafts replies and follow-ups. Nothing is sent until a person approves it in Telegram or the review app; one command pauses everything.
In your domain: replies to clients and citizens, claims correspondence, supplier communication.
of autonomy · 9 agents · 99 tools
Agents work on accounts, deals and tasks with a knowledge-graph memory. Each action class has an autonomy tier and a risk score; riskier actions wait in an approval queue with deadlines and escalation.
In your domain: case management, onboarding, back-office processes with approvals.
plus DEX, stock, forex and news feeds
AssetManiac periodically reads blockchains and decentralised exchanges alongside traditional markets and news. Analyst agents write their view; others go back to earlier calls and mark what was right and wrong.
In your domain: market surveillance, fraud and anomaly detection, compliance monitoring over large data streams.
correct "no tool call" decisions · 2,471 held-out steps
The factory behind the others: dataset building, training, frozen evaluation, regression gates, release. More than twenty models are public on Hugging Face with model cards, data cards and results.
In your domain: an agent that gets measurably better with every month of your data.
machine-verified, each with a runnable check · 34 self-built tools
Automatic discovery for agents. Dropped into an unknown system, the agent asks its own questions, commits a prediction before every experiment, and a separate judge scores the surprise. A finding enters the archive only if it replicates and its check passes. Every thread is exported as a training trajectory, which feeds straight back into the flywheel. Open source on GitHub
In your domain: an agent that maps a legacy system, database or process before anyone automates it, and turns what it learns into training data for your model.
Each action class gets a tier. Raising a tier is a human decision, logged with who made it. Quotes, prices and contracts never run on their own, whatever the tier.
| Action | Draft only | Low | Medium | High | Full |
|---|---|---|---|---|---|
| Read & analysesearch, summarise, classify | ● | ✓ | ✓ | ✓ | ✓ |
| Internal writetask, note, score, record update | ● | ● | ✓ | ✓ | ✓ |
| Routine external messagereply, reminder, confirmation | ● | ● | ● | ✓ | ✓ |
| High-value or bulkmass mailing, deal change | ● | ● | ● | ● | ✓ |
| Quote · price · contractalways a person | ● | ● | ● | ● | ● |
Every record links to the one before it. Change or delete a row and verification shows exactly where. Inputs and outputs are stored as hashes, so personal data is not copied into the log.
The EU AI Act assigns duties by role. We settle the roles in writing before work starts.
Data sources and licences, training procedure, results on frozen, versioned test sets.
Each version scored against the previous one on the same set; regressions flagged before delivery.
Graded autonomy tiers; contracts, pricing and quotes always need a person.
Agent actions and approval decisions logged with timestamps; retention you control.
Provider or deployer, risk class, transparency duties, and the documentation that goes with them.
We are not a notified body and do not certify compliance. Conformity of high-risk systems is the provider's responsibility; we supply the technical documentation and support.
Some processes need a better form, not an agent. Some need a rule engine. Some are done by a person in five minutes a week.
The one-day evaluation (€1,900) runs untuned models on fifty of your real cases and tells you, in writing, whether an agent is worth building at all. If it isn't, that is the result.
Price listAgents that do a repeated job end to end: read mixed input (e-mail, documents, voice), call your systems through tools, draft or act, and ask a person where it matters. They run on a model trained on your data, inside your infrastructure or EU-hosted.
We also own the model layer. We train the model the agent runs on, measure it on your own frozen evaluation set, and hand over the weights, so the agent keeps working without a foreign API and without your data leaving the building.
If you put the system into service under your name, you are the provider and we are the supplier. We agree in writing what we hand over (technical documentation, evaluation results, logging and oversight description, instructions for use) and classify the risk of each use case in the statement of work.
Only as far as you allow. Every action class gets an autonomy tier, from draft-only to autonomous, and contracts, prices and quotes always require a person's approval.
Pick the job your team repeats most. In a workshop we map it, find where an agent helps and where it doesn't, and tell you what it would take.