You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
We are building AIOA, an open-source safety layer for AI systems.
The idea came from a very simple problem: AI can sound confident even when it is wrong, outdated or only repeating another model. We wanted to build a system where model output is never treated automatically as truth. AIOA separates claims from evidence, keeps track of sources and contradictions, and makes sure that important actions still need human approval.
We already have a working public repository, tests, documentation and a recorded demo with a German employment-law case. We are now building the next part: a provenance-aware memory layer and modular Knowledge HATs, so the system can remember where information came from, what changed, what was contradicted and what was only generated by a model.
We are not claiming that AIOA solves hallucinations or alignment. We are building a practical system that makes AI workflows easier to inspect, test and control before they are trusted in the real world.
Our goal is to turn AIOA from a working prototype into a tested, open AI safety system.
We want to build a benchmark for outdated and unsupported model answers, improve the multi-model review process, and add a memory layer that keeps evidence, model output, contradictions and human decisions separate.
We can achieve this by expanding the existing code, running controlled tests across different models, validating the results with external reviewers, and publishing the benchmark, documentation and failures openly.
The funding will mainly be used for a reliable development computer, model API credits, cloud compute and storage.
It will also cover testing, benchmark runs, backups and some external technical review, so we can properly validate the system instead of only building it on very limited hardware.
At the moment, I am building the project alone.
I previously explored working with a colleague, but his company was focused on trading and consulting, while AIOA is a completely different type of project. We could not align the company profile and long-term direction, so we decided to continue separately.
I am currently responsible for the architecture, development, testing and documentation. When funding becomes available, I plan to bring in external technical and domain reviewers where needed.
The most likely reasons would be limited funding, weak hardware, trying to build too much as one person, or discovering that some parts of the system do not improve safety enough to justify their complexity.
The most likely outcome would be slower development or a smaller project, not a total loss. We would still publish the useful code, tests, benchmark results and also the parts that did not work. Even a negative result could help show where multi-model review, memory and provenance systems fail.
We have raised $0 in external funding during the last 12 months.
The project has been developed using my own limited money, equipment and paid AI subscriptions. One previous application passed an initial stage and is waiting for a final decision, but no funding has been confirmed or received.
There are no bids on this project.