The Short Version
In several research projects, colleagues and I have relied on commercial tools for conversational AI agents. They're good, and in many ways impressive. But they don't really cater to research designs: probing logic, follow-up strategies, semi-structured and unstructured interview protocols, participant identifiers, or the kind of pricing model that works for an unfunded PhD project.
That's why I started OASIS. It's not trying to compete with commercial platforms. It's trying to be an open-source alternative for transparent research, a survey interview platform that runs entirely inside your Docker environment, talks to whichever AI providers you choose, and treats research methodology as a first-class concern rather than an afterthought.
To be clear: the ambition here is not to match enterprise platforms with dedicated engineering teams and substantial funding behind them. They will always move faster and offer more polish. But it is important that open, transparent alternatives exist for research. Tools you can self-host, inspect, and adapt to your study design without a procurement process. That is what OASIS is for.
Read the longer version
The models are there. GPT-4o can run a solid interview. The problem is everything around the model: getting audio in and out in real time, managing sessions, storing transcripts, switching between providers without rewriting your code.
You can pay for an enterprise tool that wraps all of this up. That works until you need to switch providers, or you realise your transcripts live on someone else's server, or the pricing changes and your unfunded project is suddenly out of budget.
The other option is building it yourself — wire up WebSockets, plug in an STT engine, add a TTS service, write a session manager. Doable if you have a software team. Most research groups don't.
OASIS is the middle ground. It does the plumbing so you don't have to. It runs in Docker on your machine or your server. If you configure it to call OpenAI or Google, that's on you — OASIS itself doesn't phone home or touch your data. You can also point it at fully self-hosted models if that's what your ethics board needs.
