The Volundra AI
Our own decision system.
Our own rulebooks, our own safety architecture, our own memory,
orchestrated on leading models and operated in the EU. We do not train
a foundation model of our own, and we do not claim that we do.
Rulebook
Verified knowledge instead of gut feeling
Under every decision sits a verified reference work of the platform
rules, every statement with its source and the date it was retrieved,
anything unsourced marked as unsourced. It grows with the rule
changes of the platforms, not with the fashions of the market.
Safety architecture
Limits the client sets
Limits, risk and scope are set by the client. The system
checks itself against those limits before it acts, stops instead of
crossing them, and can be halted from the outside at any time.
Memory
Every run leaves a trace
The numbers, the decision and the reasoning in plain language are
written down, the doubts included. From that record the system learns
what carries a case and what does not, so the next run does not
start from zero.
Orchestration
Models are replaceable, we are not
For language and judgment we call leading models, chosen separately
for each task. They can be swapped out without the system losing what
it knows. Our rules, our limits and our memory stay.
Case study: the first application
The first service we took over completely runs in production today. Every
number below is measured there, in live operation. It is what coverage
looks like when it does not shrink as the load grows.
100%
of the incoming signals are reviewed every day. Whether a case produces five hundred a day or eleven thousand, each one is rated and either passes or gets blocked. The system does not sample and it does not skim.
Measured in production, not estimated
- 25
- automated checks run before the system acts, at small scale and large alike. If one fails, nothing happens.
- 9,500
- lines of verified rulebook sit underneath those checks, each statement with its source and retrieval date. It grows with every rule change, so the thousandth case starts smarter than the first.
- 7
- days a week, around the clock. The system works on the schedule of the systems it operates, not on office hours, and it does not take August off.