MVPLab built VoiceBrief.io, a platform that turns lecture notes, papers, and textbook chapters into professor-grade audio lectures with quizzes and Q&A, and took it from concept to paying customers in a matter of weeks. Claude is the backbone of the product across four production layers: OCR and structural cleanup of scanned or photographed notes, lecture script generation, spaced-repetition quiz generation, and Q&A over the student's own documents. Audio runs on Kokoro TTS self-hosted on an NVIDIA DGX Spark with Cartesia as fallback, keeping unit economics viable against an unlimited $9.99/month Pro tier. Auth, payments, file handling, error recovery, usage limits, and monitoring all shipped in the initial build. VoiceBrief reached early MRR with real paying customers, with no dev team hired and no round raised.
The problem
Students sit on piles of material: lecture notes, scanned handouts, papers, textbook chapters. Most of it never gets read. VoiceBrief's founder saw the fix clearly: upload your notes, press play, and learn on your commute or at the gym.
But "press play" hides the hard part. A robot voice reading your notes top to bottom is useless. What people actually want is what a good professor does: explain things in the right order, slow down for the hard parts, skip the filler, and check whether you understood. That's what VoiceBrief had to deliver, plus quizzes and a tutor you can ask questions.
The founder had the idea and the audience ready. What was missing was someone to build it properly, quickly, and ready to take payments from day one. That's where we came in.
How we put Claude into production
Here's the thing we told the founder early: reading text out loud is easy. Teaching it is hard. The product's real value is the judgment in between: deciding what matters in a document, what order to teach it, where to slow down, and what a student is likely to get wrong.
That judgment is the job we gave Claude. It isn't one API call buried in a chain. Claude is the product, doing four jobs in production:
1. READ THE MESSY NOTES
Photos taken at an angle, scans, two-column PDFs, tables. Claude does the OCR and puts everything back in the right reading order.
2. WRITE THE LECTURE
Claude decides what to teach, in what order, and where to slow down. This is what makes it sound like a professor.
3. WRITE THE QUIZZES
Real multiple-choice questions, with wrong answers that are actually tempting.
4. ANSWER QUESTIONS
Students ask about their own notes. Claude answers and points to the exact spot in the source.
Reading the messy notes
Real study material is never clean. It's a photo of a whiteboard, a scanned handout, a two-column paper full of tables and footnotes. Claude handles the OCR, plus the part OCR tools usually get wrong: putting the document back together the way a human would read it. Headings where they belong, columns in order, page numbers thrown away. Everything downstream depends on this step. If the input is garbled, the lecture is garbled. Moving this step to Claude is what produced the jump in audio lecture quality.
Writing the lecture
This is the heart of the product. Claude takes the cleaned-up document and writes a teaching script: what to cover first, how deep to go on each topic, where to pause and emphasize. It's the difference between a voice reading your notes and a voice teaching them. VoiceBrief calls it Teach Mode.
Writing the quizzes
Anyone can generate trivia. A good quiz question needs wrong answers that are tempting. That's what makes it test understanding instead of word-matching. Claude writes those. A spaced-repetition scheduler (SM-2, the same algorithm Anki uses) then brings each question back right before you'd forget it.
Answering questions
Students can ask anything about their own uploaded material. Claude finds the answer and cites the exact place in the source it came from, so the answer can be checked instead of just believed.
The audio layer
The voice itself runs on Kokoro TTS, which we self-host on an NVIDIA DGX Spark, with Cartesia as backup. The reason is simple math: Pro is unlimited listening for $9.99 a month. Pay a TTS API per character and your heaviest users lose you money. The product gets more expensive as it grows. Self-hosting flips that. Cost per lecture falls as usage climbs.
On top of the audio we built the feature users mention most: as the lecture plays, the exact word being spoken lights up in the original document. Word by word, not paragraph by paragraph. It took a custom layer tying audio timestamps back to the parsed document, and it's the thing people say sold them.
What "production-ready" actually meant
Getting Claude to do something impressive in a demo takes an afternoon. Getting it to do the same thing for every user, every document, every day is the real work. Around the pipeline we built the unglamorous parts: validation for broken files, retries and fallbacks when a step fails, usage limits so the free tier can't be abused, cost tracking per document, and monitoring so problems show up in a dashboard before they show up in reviews.
Auth, payments, and billing shipped in the same initial build. No "we'll add billing later." The product was collecting revenue within weeks of the first commit: free tier, $9.99 a month Pro, $99 lifetime.
The result
VoiceBrief reached early MRR with real paying users. Not a waitlist. Not beta testers. Customers.
The founder didn't hire a dev team, didn't raise a round, and didn't burn months on iteration cycles. The tech runs without babysitting: Claude handles the thinking, the self-hosted stack keeps the margins healthy, and the founder's time goes into the business.
In the founder's words
I had the idea for VoiceBrief and the audience ready. What I didn't have was someone who could actually build it at production quality, fast. MVPLab.ai took VoiceBrief from a concept to a product generating revenue in a matter of weeks.
The whole product rests on Claude, and that was their call. Turning a messy pile of lecture notes into something that sounds like a professor teaching you is not a summarization problem. It takes judgment about what matters, how to pace it, and where to land emphasis. MVPLab.ai built that on Claude end to end: the OCR pass over scanned and photographed notes, the lecture script itself, the quiz generation, and the Q&A our students run against their own documents. The jump in lecture quality when that pipeline landed is why our Pro conversions work at all.
They paired it with self-hosted TTS so our costs stay manageable as we grow, and word-level audio synchronization that users constantly mention as the feature that sold them. I never had to explain what 'production-ready' means. Auth, payments, error handling, and monitoring shipped on day one. If you have a product idea that needs serious AI infrastructure and you need it built right the first time, MVPLab.ai is who you talk to.