ToneTutor

The business

ToneTutor (https://tonetutor.tefusiang.com) teaches Mandarin the way every competitor refuses to: tones first. Learners hold real spoken conversations with an AI tutor that grades every reply's tones and — unlike the apps that just flash "wrong, try again" — explains which syllable missed, why, and how to fix it. Three free sessions, then US$5/month. No signup required to start. Revenue is live through Lemon Squeezy (merchant of record), so a solo founder in Malaysia sells globally with tax and compliance handled.

Why this business exists

I mined 2,734 recent App Store reviews of seven competitor apps — with an AI agent, in one afternoon. The pain was unambiguous: SuperChinese users spend "15 minutes on jìxù" being told they're wrong with no explanation; HelloTalk's recent reviews are 75% negative because a language-exchange app became "a dating app full of predators"; HelloChinese fans literally ask for tones to be emphasized earlier. ToneTutor's four pillars — explains why your tone is wrong, safe AI conversation with no creeps, tones-first, $5 with no signup wall — were extracted from that evidence, not guessed. The same agent re-pulls reviews weekly and rotates fresh pain points into our marketing hooks.

AI in production, making real decisions

1. The product is an AI employee. Gemini 2.5 Flash conducts every conversation, grades tones from Google Cloud Speech-to-Text transcripts, estimates the learner's HSK level, and generates personalized shadow-drills and vocabulary drills targeting that specific learner's diagnosed weaknesses. A latency-based model-routing layer sends conversational turns to Flash-Lite (~1.1s replies) and quality-critical assessments to Flash. There is no human tutor anywhere in the loop; the marginal cost of serving a session is under a cent.

2. Marketing is run by an AI agent. Customer acquisition is a daily agentic pipeline: the agent writes video hooks from the competitor-pain database, builds the videos programmatically (Python/ffmpeg — no editor, no designer), voices them with Google Cloud TTS in the same bilingual voice persona users hear in the app, schedules them to YouTube via API at data-chosen times, and replies to viewer comments in the commenter's own language. When abstract grammar-lesson videos flopped (2–8 views) while embarrassment-stakes hooks pulled 300–415, the agent diagnosed it from the numbers and reset the content strategy itself. The human approved it in minutes.

3. Analytics close the loop. A first-party event funnel (test_started → test_completed → card_rendered → paywall_hit → subscribe) is stored in Firestore and read daily by the agent. Every posting-cadence change, pricing experiment, and feature priority traces to those numbers.

The human's job has shrunk to: approve the day's plan, handle the one thing platforms require a human for, and be the bilingual native-speaker taste-check on voice quality.

Built on Google Cloud

Cloud Run (scale-to-zero, asia-southeast1) hosts the FastAPI app; Vertex AI Gemini 2.5 Flash / Flash-Lite power tutoring, grading, and drill generation; Cloud Speech-to-Text (LINEAR16 PCM — including a Web-Audio workaround so iOS mics work) and Cloud Text-to-Speech (Chirp3-HD) give the tutor ears and a voice; Firestore stores users, subscriptions, and the analytics funnel. One gcloud run deploy ships the whole business.

Traction

We would rather show small real numbers than large vague ones.

• Two arms-length paying customers — strangers in Germany and Canada, who each bought the $19 founding-lifetime licence on 14 and 15 August 2026, within nine hours of each other. A third order is the founder buying his own product to verify the live billing path, and is disclosed separately as related-party revenue.
• A 27% paywall conversion — of everyone who reached the paywall, better than one in four bought. 55% of the people who start a tone test finish it, with no signup required.
• $0 marketing spend. 125 Shorts on @tonetutor, all written, voiced, rendered and scheduled by an agent. Customer acquisition cost is $0.
• Revenue, expenses and the full profit-and-loss statement are itemised in the judging materials.

Two sales is two sales. That is not product-market fit and we are not going to dress it up as one. What it does prove is that the entire path works end to end: a stranger on the other side of the world can find the app through AI-made video, use it without signing up, hit the paywall, and pay — with tax collected and remitted correctly in their own jurisdiction. The constraint was never the product or the funnel. The constraint is throughput into it.

Business viability

• Price: US$5/month — deliberately under the "subscription fatigue" threshold our review-mining surfaced (HelloTalk VIP is $13/mo). Cheaper than one bubble tea.
• Unit economics: COGS ≈ $0.02–0.05/user/month at current usage. Gross margin > 95%. CAC ≈ $0: all acquisition is AI-generated organic short-form content.
• Billing: live since June 24, 2026 — real checkout, verified webhooks (subscription_created / order_created → Firestore Pro), 3-free-sessions paywall.
• Market: tens of millions of active Mandarin learners globally (HSK candidates, heritage speakers, scholarship applicants); tones are the universally-acknowledged hardest part and the least served.

Challenges

Making an AI grade speech honestly (STT noise vs real learner errors), keeping conversational latency ~1s on a $0 infra budget, and — the business challenge — converting short-form views into paid users as a solo founder with no ad spend. Each fix in that funnel was made by the AI reading its own data.

What's next

Real-time F0 pitch-contour analysis (grading the actual tone curve, not just the transcript), founding-member pricing to accelerate the first cohort, and expanding the review-mining → hook-generation loop into a fully autonomous weekly marketing sprint. The 90-day story: one person + Gemini + Google Cloud shipped a revenue-generating school where the AI is the teacher, the marketer, and the analyst.

Built With

• cloud-run
• fastapi
• ffmpeg
• firestore
• gemini
• google-cloud
• javascript
• lemon-squeezy
• python
• speech-to-text
• text-to-speech
• vertex-ai