Company
Built for work that has to be defended.
Finance is not judged on whether an answer sounds right. It is judged on whether the analysis behind it can be traced, attributed and defended — to a credit committee, to model risk, to an examiner, sometimes years later. That is the bar the work has always had to clear, and it did not move when general-purpose AI arrived.
General-purpose assistants cannot clear it. They produce fluent output detached from its sources, blur the line between what a person wrote and what a model wrote, and treat every action as equally reversible. That is tolerable when the output is a first draft of an email. It is not tolerable when the output is a credit memo, a suspicious activity narrative or an investment brief.
Versidon exists because the answer is not a better chat window. It is a platform built around the three things the work actually requires — evidence, attribution and control — with the AI arranged to serve them rather than the other way round.
The thesis
Analysis you cannot trace is analysis you cannot sign.
So the agent works inside the files a finance team already keeps, every figure it writes carries the filing and the page it came from, everything it drafts is marked and reversible until a person accepts it, and nothing irreversible happens without the holder of the entitlement. Those are not features bolted onto a chatbot. They are the shape of the platform.
What that means in the product
- 01
Every number carries its source.
Click any figure in a brief, a memo or a case narrative and Versidon shows you the passage it came from — the filing, the page, the line. Analysis you cannot trace is analysis you cannot sign.
- 02
You always know who wrote what.
The agent drafts into the same file your team works in, and every block it writes is marked, attributed and reversible until a person accepts it. The audit log records the actor, the entity and the timestamp for both.
- 03
Nothing irreversible happens on its own.
Releasing a payment, filing a report, approving a facility — Versidon prepares the pack and names the residual risk, then stops. The decision stays with the person who holds the entitlement for it.
Operating principles
Three rules we would rather lose a deal than break.
- If it cannot be traced, it does not ship
- A feature that produces a figure without a path back to its source is not finished. When the agent cannot ground a claim, the correct behaviour is to leave the field unresolved and say why — not to fill it in convincingly.
- The person keeps the decision
- We do not build automation that completes an irreversible step. Releasing a payment, filing a report, approving a facility: the platform prepares the pack, names the residual risk and stops at the entitlement holder. That is a design constraint, not a setting.
- We publish the limits with the capability
- Documentation states where the platform performs worst and under what conditions an evaluation suite fails. A reviewer who only ever hears what a system does well has not been given enough to review.
And the other side of it
- We do not train models on customer data, and no customer data crosses into a shared model.
- We do not ship an agent that can approve its own work, or inherit a person's approval rights.
- We do not present model-generated figures without the passage they came from.
- We do not claim certifications, customers or results we cannot evidence on request.
Careers
We hire people who have had to defend an answer.
Engineers who care more about retrieval and evaluation than about demos, and people who have worked in credit, research, compliance or operations inside a regulated institution and know exactly where this kind of software usually disappoints. We are not running a public listing of open roles. If that describes you, write to us and say which part of the work you would want to own.
Get in touchContact
One route in, answered by a person.
Evaluations, security and model-risk reviews, partnership and press all come through the same form, and a person reads it. Tell us who you are and what you need; we reply within two business days. If you want the briefing packs — the implementation guide, the model-risk documentation, the security review pack, the evaluation methodology — say so and we will send them.