AI
AI that shows its work.
Rules where rules are enough, models where they help, and never a model deciding on money or protection. Each feature says what it is and what it does not do.
Risk center
An alert you can explain.
Every alert names the rule that fired and the threshold behind it, so the team sees why a day stands out.
Stake jump
Flagged because this day's stakes are more than three times those of the previous day: a rule with a threshold the operator sets, not a learned model.
- In production
Risk center
Per operator, 27 alert types check deposits, withdrawals, stakes, sessions, shared IPs, RTP and AML signals against thresholds the operator sets, most of them every five minutes. Alerts can go to Slack or email.
Details
These are rules with per-operator thresholds, plus two simple statistics: the cadence of rounds and observed against declared RTP. It is not machine learning and it does not learn. - In production
Reconciliation on a schedule
Wallets against the ledger every six hours with 18 rules. Open deposits against the gateway every minute, withdrawals and refunds every five. Provider callbacks every six hours, the audit chain every hour.
Details
It finds and reports; it never corrects money on its own. Every mismatch goes to a person. Gateway checks run against test providers today. - In production
Recovery jobs
A paid deposit whose webhook was lost is recovered by asking the gateway, and so are unconfirmed withdrawals and refunds. Outbound webhooks get up to five attempts over about two and a half hours, and a scheduled job that misses three runs raises an alert.
Details
Orphan game rounds are not rolled back automatically: they go to review. - In production
Business health check
Nine invariants, on demand from the console overview: every transaction belongs to the right wallet, balances match the ledger, locked funds match pending withdrawals, no balance is negative, and how old the withdrawal queue is.
Details
It reports what it finds, with the wallets affected. It does not block or repair. - In production
Scores explained factor by factor
Churn risk from six weighted factors in five bands, fraud by rule points, value on a log scale and RFM by quintiles. Each score shows what moved it.
Details
No model is trained. A player with a responsible gaming signal keeps the score, but no retention surface acts on it. - In pilot
Generative AI on a model set per deployment
One provider layer for every generative feature: by default an OpenAI-compatible model server set for each deployment, or a hosted model only when it is configured with its key. There is no automatic fallback to a paid service.
Details
Configuration is platform-wide; each operator has a spending cap. The model server's availability in production has not been verified. - In pilot
Bet builder in plain words
A player describes a bet in plain language. The model only picks from the listed options; the server sets the price and the margin.
Details
It needs the model server and a live odds feed, both in pilot. Same-match combinations are not supported, so its picks are placed as singles. - In pilot
Support assistant
A bot that answers from the operator's knowledge base and the player's own data, translation of conversations for agents, and a three-line summary when the bot hands over. Card numbers, IBANs and identity documents are masked before anything reaches the model.
Details
The bot is off by default. Translation is on by default and shows the original message when the model does not answer. - In pilot
Drafts for site texts
In the console, the model drafts translations of an operator's site texts and marks them for review. Nothing is published without a person.
Details
Drafts only. Translation of promotions, SEO and pop-ups exists in the API without a screen yet. - Coming soon
Computer vision for KYC
Our own engine reads identity documents, compares the face with the document and checks liveness, with no network access. It never approves a case on its own.
Details
Not deployed yet, and its liveness check is not certified. - Coming soon
A decision layer
Recommendations per player under the operator's policies, run in shadow first and measured before anyone acts on them.
Details
Not built. The event log it needs is deployed and switched off.
Principles
Rules where rules are enough
Monitoring, reconciliation and scores are rules with thresholds an operator can read and change. We call them rules, because that is what they are, and because a compliance team has to be able to explain every alert.
A model server per deployment
Generative features run on a model server chosen per deployment, in pilot today. A hosted model is used only when it is configured with its own key, and nothing falls back to a paid service on its own. Each operator has a spending cap.
A person decides
Models draft and suggest. They never publish a text, approve an identity check or move money. Automatic KYC approval is refused for operators with real money, and money corrections above a threshold take two people.
Masked before the model
Card numbers, IBANs and identity documents are masked as messages arrive by chat, email or phone, before they are stored, translated or seen by a model. The original is sealed, and only compliance can open it, from a responsible gaming alert.
Contact
Tell us what you are building.
Markets, brands, verticals. We show you the platform with your case in mind.