OliverWales

I'm not sure where I first heard of rubber duck debugging, but in my experience there is rarely a physical rubber duck at all. It's more of a "can I rubber duck you on something for a minute?" "Yea, sure!" "<starts explaining> ... ah I've figured it out cheers".

If you put it in a sun-synchronous orbit then it will always be in sunlight. Such an orbit would be in LEO.

Thanks for the feedback! Most people have been using free online templates to create docs (or random unrelated PDFs) but we'll consider adding sample documents next time.

If we fix business banking onboarding there will be no need for this!

It is not possible to give the reason for offboarding as it may tip off fraudsters on how they can evade detection next time. This is standard across the industry.

We're predominately selling to US neobanks and fintechs, right now focussing on the smaller ones that are less tied into their existing solution.

KYB is vastly more complex than KYC because there are more things that need to be checked, and for each one a large range of acceptable proof documents. There are also many more edge cases e.g. change of business name or structure, foreign beneficial owners etc. that make an end-to-end solution tricky.

There are other players in this space but they tend to be tools to assist human analysts. We want to fully automate the low risk cases so that they never require human intervention.

The compliance analysts that we have spoken too feel stretched too thin with a huge backlog of cases. We free them up from having to do menial document verification for low risk companies so they can spend more time on the more interesting high risk work.

Thanks for playing, this is useful info!

There's a big difference between the nature of fraudulent documents submitted to our customers today, and the ones that one might create knowing that the system is build with LLMs - we've mostly optimised for the former so far.

I'm also glad to see it took a number of attempts - security through obscurity is not something we want to rely on but the real system requires your identity and will offboard you without explanation at the first hint of misbehaviour.

We'll continue to improve our system to be more vigilant and make fewer assumptions.

Thanks!

Try it on the demo, let us know how it goes!

Yes, I would definitely put Arva in that category.

For now, we limit the amount of decisioning that is made by an LLM and make as much of the business logic as we can concrete in code. It's mostly used to extract information from documents, crawl websites and identify specific fraud signals.

We use Gemini as they offer stronger guarantees around not training on your data ve e.g. OpenAI. We are also looking at self-hosting open source models in the future.

Our agent implements every step of the procedure that a human compliance analyst would follow, step by step. As a result, every individual check that makes up the overall decision is documented with the reasoning and evidence the agent used.

You can also choose to bucket your customers by risk level and require manual review for higher risk customers (e.g. certain business activities like crypto or gambling, complex ownership structures or foreign UBOs). Ultimately the level of automation vs review is up to the risk appetite of the customer, as they are the one that has to answer to the regulator.