8. The Next 90 Days
The plan assumes about 15 hours a week. With full-time effort, the same steps fit in about 60 days.
The single most important thing
Get 50 people to pay for 30 days of work-scene speaking practice on WhatsApp before building the map.
- Audience: Hindi- or Bengali-medium learners aged 18–30 heading into work. They are job seekers preparing for interviews, or early-career BPO, retail, hospitality and office staff who understand English but freeze when speaking.
- Job: "Get me ready to speak in my interview (or on my calls) without freezing, at my level, and tell me in my own language exactly what to say better."
- How: run a concierge version.
- A form (Tally or Typeform) collects the learner's goal.
- A 5-minute voice-note placement ("introduce yourself; describe this photo; answer these 3 questions").
- A daily scene prompt on WhatsApp: "You're at the hotel desk; a guest says…".
- The learner replies with voice notes.
- An LLM drafts the feedback in Hindi or Bengali plus a better version. You check it, and the trainer spot-checks it.
- A weekly score and a before/after clip.
- Charge from day one: ₹149–199 for 30 days.
Week by week
| Weeks | Goal | Exit criteria |
|---|---|---|
| 1–2 | Scenes, rubric and the voice test. With the trainer, write 5 work scenes (interview intro, customer complaint call, hotel check-in, shop-floor query, team stand-up) at 3 levels. Define the level scale and the feedback rubric. Test 2 voice stacks (Sarvam STT + small LLM + cached TTS, and Gemini Flash Live) on 20 real learner recordings. | Rubric agreed with the trainer. STT accuracy on Indian-English learner speech good enough that feedback is right ≥90% of the time (trainer check). Cost per voice minute ≤₹1. |
| 3–4 | Pre-sell. Post 20 short videos ("say this in an interview, not this") in Hindi or Bengali. Share in 5 job-seeker communities (WhatsApp, Telegram, college groups). Run the concierge. Charge for it. | 50 paying. Below 20 means the hook is wrong. Test customer-call-first or interview-first before anything else. |
| 5–10 | Thin v1 (web app / PWA + WhatsApp): speaking placement → 2D map with 5 places → scene engine ("scripted spine, live edges", cached NPC audio) → per-turn feedback in Hindi or Bengali → weekly speaking score with before/after clip → share card. 18+ gate, AI disclosure and audio deletion from day one. Razorpay UPI Autopay with pre-debit notices. | Concierge users move over. They practise ≥4 days a week. |
| 11–12 | Measure and decide. | 150+ payers, ≥50% of monthly payers renew for month 2, median speaking score +10% in 30 days, ≥20% of new users from shares or referrals, voice and model cost below 20% of revenue. |
What to measure from day one
- Activation: finishes the speaking placement and one work scene (target ≥50% of signups).
- Aha: "Does your level feel right?" (target ≥70% yes). "Did the feedback help?" rated 1–5 (target median ≥4).
- Habit: practice on ≥4 days a week among payers; day-7 return ≥25%.
- Progress: 30-day speaking-score change (target median +10%).
- Money: free-to-paid ≥3% within 14 days, month-2 renewal ≥50%, refunds <3%.
- Cost: voice and model cost per payer ≤₹40 a month.
- Loop: share of new users arriving via a score card or scene challenge.
Kill criteria
Stop and change course if either of these happens:
- fewer than 20 paying by day 30;
- month-2 renewal below 35% by day 90.
Either one means the hook or audience is wrong. Change the entry scene (interview, customer call or hotel desk) or the audience (job seekers vs employed staff; Hindi-medium vs Bengali-medium). Don't respond by adding Hindi and Bengali target-language courses, a casual track or new game mechanics.
The one thing to stop
Stop designing the city. The map, the avatars and the three-language curriculum are the most enjoyable parts to build and the least likely to decide whether anyone pays twice. The WhatsApp concierge tests the paid core in a week. The map can be a 2D illustration with 5 tap zones, built once the scenes prove themselves.