Chain Academy
A Voice AI learning companion that converts blockchain and Web3 educational content into interactive spoken conversations. Instead of reading a course, learners talk to it — asking questions, requesting examples, and getting answers grounded in the source material.
The problem
Educational content in Web3 is dense, fast-moving, and mostly written for people who already understand it. Newcomers hit a wall: long-form articles, dense whitepapers, and video courses that assume prior knowledge and offer no way to ask "wait, what does that mean?" in the moment. Passive reading and watching produce low retention, and the drop-off between "started a course" and "actually understood it" is enormous. The gap isn't a lack of content — it's a lack of interactivity. Learners can't interrogate the material, can't get it re-explained at their level, and can't turn a static lesson into a two-way conversation. The knowledge exists; the conversation that unlocks it doesn't.
Key challenges
Educational conversation is harder than transactional voice AI. The system can't just retrieve a fact and read it back — it has to explain, check understanding, rephrase when the learner is confused, and stay grounded in the actual curriculum instead of hallucinating plausible-sounding blockchain nonsense. Accuracy matters more here than almost anywhere: a wrong explanation of how a consensus mechanism or a smart contract works actively harms the learner. Voice adds its own constraints. Spoken answers have to be concise — nobody wants a 400-word monologue read aloud — while still being complete. The AI has to handle interruptions ("no, go back"), track where the learner is in a lesson, and remember what's already been covered so it doesn't re-explain the basics to someone three modules in.
What we built
Chain Academy ingests educational content — articles, course modules, transcripts, and reference docs — and turns it into a grounded knowledge base the voice AI speaks from. Every spoken answer is retrieved from and cited against the source curriculum, so explanations stay accurate to what the material actually teaches. Learners talk to it like a tutor. They ask a question, get a concise spoken explanation, and can follow up — "give me an example," "explain that more simply," "how does this relate to the last lesson." The AI tracks the learner's position in the curriculum, adapts depth to their apparent level, and keeps a running memory of the session so the conversation builds instead of resetting. Built on the same production Voice AI stack behind CAIDEN — LiveKit for real-time audio, Deepgram for transcription, ElevenLabs for natural speech — tuned for the back-and-forth cadence of teaching rather than transactions.
Our approach
- 1
Content ingestion into a grounded knowledge base
Educational material — articles, module transcripts, reference docs — is chunked, embedded, and indexed so every spoken answer is retrieved from the actual curriculum. Grounding is the difference between a tutor and a confident guesser.
- 2
Retrieval-grounded spoken explanations
Answers are generated from retrieved source passages and shaped for voice — concise, complete, and phrased for listening rather than reading. The model explains; it doesn't recite a wall of text.
- 3
Curriculum-aware session memory
The system tracks where the learner is, what's already been covered, and their apparent level, so it stops re-explaining basics and adapts depth as the conversation progresses.
- 4
Real-time conversational voice loop
LiveKit + Deepgram + ElevenLabs deliver low-latency, interruptible speech tuned for teaching — the learner can cut in, ask to go back, or request a simpler explanation mid-answer.
Key architectural decisions
Ground every answer in the source curriculum
In education, a fluent wrong answer is worse than no answer. Retrieval grounding keeps explanations tied to what the material actually teaches instead of the model's priors about blockchain.
Optimize responses for listening, not reading
Voice answers have to be short enough to hold attention but complete enough to teach. Tuning length and phrasing for the ear was a first-class design constraint, not an afterthought.
Track curriculum position as session state
A tutor who forgets what you've covered is exhausting. Curriculum-aware memory lets the AI build on prior turns and adapt depth to the learner's progress.
Reuse the CAIDEN Voice AI stack
The real-time audio, transcription, and TTS infrastructure was already production-hardened. Building on it meant the effort went into the teaching experience, not re-solving low-latency voice.
Results
- Educational content converted into interactive spoken conversations
- Retrieval-grounded answers cited against the source curriculum
- Curriculum-aware memory that adapts depth to the learner's progress
- Concise, listenable explanations tuned for voice rather than text
- Interruptible, low-latency conversation loop built on the CAIDEN stack
- Follow-up questions, examples, and re-explanations on demand
- A learning experience that turns passive content into two-way dialogue
Impact
Chain Academy reframes educational content as a conversation instead of a document. Learners stop scrolling and start asking — and because every answer is grounded in the curriculum, they get explanations they can trust at the exact moment they're confused. It's the difference between reading about Web3 and having someone patient explain it to you, on demand, at your own pace.
Tech stack
Want a case study like this?
30 minutes. We scope the real problem and figure out what to build.
Book a call
