Project summary

Everybody uses LLMs accepting the inherent risk that they hallucinate. The common way to detect this is to use a second language model in the loop to judge the first one. That is expensive, it is slow, and no one can really audit why the judge decided what it decided. We can even say its a contradiction: the judge/s can hallucinate too.

I use the geometry of embeddings (where text are represented in numbers) to score whether an answer actually relied on its source, with no second model involved. Source can be a document or the reality itself. This is why I call grounding methods. The score is deterministic, so the same input always gives the same result. It is cheap enough to run on every response (important for enterprise deployment). And it is simple enough that an outside auditor can understand what's happening.

It also has a limit: when a wrong answer uses the same words as the right one, a wrong number, a wrong name, the geometry cannot detect the difference. I call this the register wall. This is a real limitation that makes the method 70% reliable as a whole. Used as first step filter its reliability is around 80-90%.

I am asking for support to make this more reliable and reproducible.

What are this project's goals? How will you achieve them?

My project is open source (https://github.com/groundlens-dev) and the idea is to create a trustworthy detection tool accesible from personal users to large deployments in regulated industries like finance, legal or healtcare. To achive this goal I need to work in testing large amounts of data from different contexts with these methods. This tests will help to imorove the methods, validate them and publish the results in order to achice the most important goal: trust.

How will this funding be used?

I need to run experiments implying several days of GPU use, generate datasets and annotate them by different individuals (not just me).

Who is on your team? What's your track record on similar projects?

The team is just me for now. I collaborate with other researchers on some topics. My track record is that I created Groundlens and have worked on the project for a year and a half. I have also advised a few consulting clients, which gave me a clear picture of the industry's trends and bottlenecks.

What are the most likely causes and outcomes if this project fails?

As said, we can trust 70% in the whole groundlens approach. As first filter around 80%. The most likely way the project fails is that these numbers do not improve. If that happens, the outcome is that the values stay where they are today. That is still a useful first filter, and the dataset and the code stay public for anyone to build on. But real trust needs a value very close to 99%.

How much money have you raised in the last 12 months, and from where?

I haven't received any funding. I have paid all the bills.