Kensink Labs
EDTECH · K-12 · TEACH-BACK AI TUTORINGLearnBuddy AI· BD

LearnBuddy AI.
An AI tutor that listens, then asks the kid to teach it back.

Cloudflare-edge · multi-agent · BN + EN.

5-step
Feynman teach-back loop, production
MULTI-AGENT VOICE-FIRST TUTORING PLATFORMEDTECH · K-12 · TEACH-BACK AI TUTORING

LearnBuddy AI: An AI tutor that listens, then asks the kid to teach it back. Cloudflare-edge · multi-agent · BN + EN.

LearnBuddy AI is a cross-platform after-school learning app for children 6 to 14 built around the Feynman teach-back method. Kids learn a concept, explain it back to Buddy by voice, get a feedback report in under five seconds, and refine until mastery. Voice-first Explain It loop with Whisper, homework scanning via Vision OCR, a multi-agent supervisor + sub-agent prompt system running on Cloudflare AI Gateway, and a parent dashboard that finally answers 'did my kid actually learn anything tonight?' Fully edge-native, bilingual day one, eval-gated before every release.

5-step
Feynman teach-back loop, production
<5s
Voice explanation to feedback report
0
Servers to operate (Cloudflare-only)
BN + EN
Bilingual voice + UI from day 1
01 · THE PROBLEM

Where they were stuck.

Most kid-learning apps are drill machines or video libraries with no feedback loop. They confirm that a child pressed buttons, not that the child understood anything, and they hand parents a sticker chart while real understanding stays invisible. Existing AI tools were not built for kid safety, can not reason between Bangla and English in the same session without dropping context, and have no way to grade an explanation a six-year-old just spoke aloud. Private tutors solve the feedback problem but cost more than most families can sustain and vary in quality week to week. The market needed one app that combines a guided teach-back loop, kid-safe gamified UX, multi-subject AI tutoring, and a parent dashboard, on a stack that scales without burning a year on infrastructure.

02 · OUR APPROACH

How we built it.

  • 01Locked the product on day one as a teach-back loop, not a quiz app. Every screen and every AI call serves the five-step Feynman cycle (Learn It → Explain It → AI Feedback → Refine It → Master It). Anything that does not move a child through the loop gets cut. The loop is the moat.
  • 02Cloudflare-only edge stack: Hono on Workers for the API, D1 for transactional data, R2 for audio + homework-photo storage, KV for prompt cache and session state, Queues for async voice processing and feedback generation. Sub-200ms p95 globally, zero servers to operate.
  • 03Voice-first Explain It pipeline: Whisper STT with kid-safe profanity + PII redaction, ElevenLabs + Google TTS for Buddy's voice with grade-appropriate vocabulary, full transcript-to-feedback round-trip under five seconds so the child does not lose focus while waiting.
  • 04Homework solver: Google Vision OCR scans a phone photo of a worksheet, the system parses each problem into structured form, then a guided solving session walks the child through (not at) each one. Every session writes to a homework log the parent reviews.
  • 05Multi-agent prompt system on VoltAgent + Cloudflare AI Gateway: Buddy is a supervisor that routes to specialist sub-agents (math tutor, English tutor, homework helper, quiz generator). Prompts live in D1 with KV cache so non-engineers can iterate language and pedagogy without a deploy.
  • 06Kid-safe by default: input + output guardrails block PII, inappropriate content, off-topic prompts, and any attempt to socialize the AI as a friend. Buddy is a tutor, never a chatbot. Parents toggle subject-level controls, daily time limits, and review weekly reports gated behind a Parent-Mode PIN.
  • 07Bilingual BN + EN from day one. Voice prompts, AI feedback, parent reports, and the UI all switch language without re-prompting the model mid-session. The Bangla TTS voice is tuned for grade-school clarity, not adult news-anchor cadence.
  • 08Eval suite before launch: hundreds of labeled teach-back transcripts per grade and subject for measuring feedback accuracy, plus red-team prompt sets for safety. The feedback engine has to pass the eval bar in CI before any release ships.
The whole product was built around one constraint: a six-year-old has to be able to explain ratios out loud and get a useful report back in under five seconds. Every architecture decision after that became simpler.
Lab engineering
Kensink Labs
[TECH STACK]
  • TypeScript
  • Flutter
  • Next.js
  • Hono
  • Cloudflare Workers
  • Cloudflare D1
  • Cloudflare R2
  • Cloudflare KV
  • Cloudflare Queues
  • Cloudflare AI Gateway
  • VoltAgent
  • Firebase Auth
  • Whisper STT
  • ElevenLabs TTS
  • Google Vision OCR
[ENGAGEMENT]
Duration8-week MVP sprint, ongoing v1 evolution
ClientLearnBuddy AI
ShapeMULTI-AGENT VOICE-FIRST TUTORING PLATFORM
HandoffFull ownership · 90-day warranty
START YOUR OWN PROJECT

Bring a real problem.
We’ll bring code on day one.

← All casesCASE · LEARNBUDDY