AI Portfolio
Ask Rebel
A voice and text assistant, grounded in this site
Try it
Ask it anything about Digital Rebel
This is live, and you can use it right now. Type a question, or tap the mic and talk to it. Ask what I do, how a service works, or which case is closest to your problem.
It also drops in clickable links to the right page, so you can jump straight to whatever it mentions instead of hunting for it.
The whole point is that it only knows this site. It isn't a general chatbot wearing my logo. It answers from the actual services, cases and portfolio content published here, and when it doesn't know something, it says so instead of inventing an answer.
It's still in testing, so you may catch it being more careful than it needs to be. That's the honest status, and there's more on what's changing below.
Under the hood
Grounded in the site, not in guesswork
I built it the way I'd build any agent I trust: thin model, strong context, clear guardrails.
The knowledge isn't a separate copy that goes stale. It reads the same content modules that render the live pages, so when I update a service or publish a new case, the assistant knows about it without me retraining anything. If it isn't on the site, it isn't in the answer. That single rule is what keeps it honest.
There's no vector database and no retrieval step here, and that's deliberate. The site is small enough that the whole thing fits in the model's context at once, so reaching for the heavy RAG machinery would have been solving a problem I don't have. The simplest setup that works, and not every AI feature needs more than that.
For voice it speaks back in a natural ElevenLabs voice, not the robotic browser default it started with. Every question gets logged to Notion so I can see what people actually want to know, and that log is privacy-safe: the question and answer, nothing that identifies you.
The build
- Next.js on the live site, one streaming API route
- Claude as the model, chosen for low latency so voice feels like a conversation, not a wait
- The whole site content passed to the model as context, prompt-cached for speed, no vector database
- Voice out via ElevenLabs for natural speech, voice in through the browser
- Question log in Notion, no personal data stored
The model is the easy part. The work that makes it usable is the grounding and the guardrails, the same lesson as every other system on this page.