A companion with a tutor’s purpose.
I propose building Morrow as an English-led language learning app. The first test is whether a consistent character helps people return to practice that produces useful, demonstrable language skills.
- One product, two jobsThe companion creates familiarity and a reason to return. The tutor chooses the next task, explains mistakes, gives a retry, and checks later whether the learning stuck.
- English is the bridgeOnboarding, instructions, grammar explanations, feedback and journals are in English. Learners hear, read, say and gradually write Korean; Japanese follows on the same principles. Tamil belongs in acquisition and community communication.
- Start with a useful sliceOne original companion, one small world and a reviewed seven-day Korean experience. Extend to a four-week beginner pilot: 20 core lessons, review sessions and four checkpoints.
- Make progress visibleShow what the learner can do without help, what needs practice and what comes next. The world celebrates participation; skill badges require evidence from independent attempts.
- Reuse proven foundationsThe existing project supplies world navigation, character reactions, voice activity prototypes and memory plumbing. Curriculum, Korean speech, durable progress, review scheduling and assessment remain real build work.
- Fund a bounded testRetain the proposed $4,000 monthly envelope for November–February, up to $16,000, subject to Olli’s agreement and monthly gates. This funds a pilot and evidence for a commercial decision.
What the name means: Morrow points to tomorrow’s small improvement and works across both languages and future markets. It is a working product name; brand, domain and trademark availability have not been cleared. Learners can give their companion a nickname.
Make the relationship serve the learning.
Tolan presents an everyday AI companion with conversational memory and expressive, nonhuman characters. Its design account describes a move toward a peer-like presence. These are useful design references; their effect on language learning is our hypothesis to test. [1] [2]
Keep one voice, personality and expressive character across lessons. The learner knows who is greeting them.
Remember the learner’s goal and chosen interests. Turn yesterday’s difficulty into today’s useful practice.
A short English journal recalls an actual learning moment and previews a small shared adventure.
Frames from the supplied Tolan reference film, used to discuss the direction. Morrow is an independent concept; these are not proposed Morrow character designs.
“I remember your goal.”
Personal interests, warmth, a consistent voice, small callbacks and a life in the world. It makes a practice session feel personal.
“Here is your next step.”
Reviewed objectives, prerequisites, targeted feedback, retries, delayed review and independent checks. It decides the learning sequence.
“Look what you can do.”
A café becomes a request task; the station becomes a listening task. Visible changes mark milestones the learner understands.
An original guide with a clear personality
- Direction
- A curious little traveller with a lantern, an expressive face and a calm, playful manner. Final character design is a build deliverable.
- Competence
- Knows the lesson language and teaches accurately. Playful “teach me back” moments are retrieval exercises, clearly framed as practice.
- Voice
- Warm English coaching and native-reviewed target-language examples. Test whether one voice can do both well before locking the voice.
- Daily life
- Prewritten small stories, hobbies and world activities, with short live responses grounded in the learner’s session.
The relationship has useful boundaries
Return prompts offer a manageable next step. Missing a day does not hurt the character or erase the world. The companion welcomes mistakes, respects a pause and never treats affection as a reward for spending.
Personal memories are visible, editable and optional. The journal reports recorded events; it does not invent progress or private knowledge.
“You wanted to order without switching to English. Yesterday you needed the phrase card. Let’s try a new drink today.”
Every session has a teaching decision.
A lesson needs an observable outcome, enough support to attempt it, useful correction and a later check. The language model can make the dialogue natural; the lesson engine owns the sequence and the progress record.
- Recall · 1 min
Try two due items before seeing an answer. Find what needs another pass.
- Understand · 2 min
Hear a short example. Get one plain English explanation and its social context.
- Notice · 2 min
Choose what was heard, connect sound to script and distinguish meanings.
- Produce · 3 min
Speak with decreasing support. Get one useful correction and try again.
- Transfer · 3 min
Use taught material with a different item or situation, without a sentence model.
- Remember · 1 min
Save evidence, schedule review and show the next small goal.
The 12-minute structure is a pilot design, adjustable for the learner. Spacing and retrieval research support revisiting material; they do not establish this exact session length or guarantee fluency. [14] [15]
Two kinds of memory
Skill, task version, response, hints, attempts, assessment certainty, last independent success and next review. Track listening, script, vocabulary recall, form/register and spoken task completion separately.
Preferred name, interests, motivation and optional personal details. These can shape a scenario or callback. They cannot establish that a language skill has been learned.
Adapt the next task to the actual gap
- Recognizes, but cannot recall: remove answer choices, offer a smaller cue, then ask again later.
- Can read, but cannot follow audio: use another recording and a focused listening contrast.
- Meaning works, register does not: explain the situation, prompt self-repair, then model if needed.
- Several attempts fail: simplify the English explanation and reduce the task before retrying.
- Audio is uncertain: ask for a retry or offer typing. Save the attempt as uncertain; a typed response can show wording, not spoken ability.
| Skill state | Evidence needed | What happens next |
|---|---|---|
| New | No prior attempt for this objective. | Short model, meaning and guided practice. |
| Practising | Success with a cue, repetition or immediate retry. | Fade support; schedule another retrieval attempt. |
| Independent | Proposed rule: unaided success on two separate sessions and a new-context task. | Show a specific “can do” achievement, then revisit it later. |
| Review due | Time elapsed or a later attempt exposed a gap. | Retrieve again; shorten the interval after difficulty. |
Initial review intervals: 1, 3, 7 and 14 days, adjusted after difficulty. These are configurable product rules for the pilot, not a universal scientific schedule. [15]
One café. One useful new skill.
Explore a scripted lesson flow from the learner’s side and the tutor’s side. Both course examples use English explanations; the target language changes.
Ask politely for a drink.
“Let’s get a coffee. I’ll help with the phrase, then you can try a different drink.”
In the app: hear the example, practise aloud and receive feedback.
Scripted concept walkthrough. Buttons simulate lesson decisions; there is no microphone, live AI, pronunciation assessment or saved learner progress in this brief. Final course content requires qualified language review.
The first visit, before this café lesson
Choose Korean, confirm comfort with English, pick a motivation and meet the companion. Offer a short placement check or a beginner start. Teach one useful greeting and one sound–script connection; finish with a successful small attempt and a clear next step.
The following visit
The companion recalls the learner’s goal. Due review comes first, followed by one new task. A two-minute “busy day” review keeps the habit manageable. Longer sessions can include bounded live talk using familiar material.
Useful speech, literacy and review grow together.
Korean comes first to focus curriculum work and the initial audience test. Use King Sejong’s introductory and standard curriculum as references for scope and sequencing; author our own reviewed lessons. [10] [11]
| Pilot stage | Practical outcomes | Knowledge developed alongside | Independent check |
|---|---|---|---|
| Week 1 · Arrive | Greet, introduce yourself, ask for repetition. | Hangul sounds and syllable blocks; a small set of polite chunks. | A new greeting exchange and decoding taught sound patterns. |
| Week 2 · Explore | Request familiar items; express a simple preference. | Menu words, request forms, sound–script matching. | Order another taught item without a phrase card. |
| Week 3 · Connect | Describe a simple routine and answer basic questions. | Selected common verbs and time expressions; listening contrasts. | Understand a new recording and give a relevant response. |
| Week 4 · Put it together | Combine introductions, preferences and a simple plan. | Cumulative review, short written production and repair strategies. | A fresh guided scenario, delayed recall and reading tasks. |
Authoring plan: five core lessons plus review/checkpoint opportunities each week. Pacing adapts to evidence; these are not deadlines for every learner. This pilot does not promise fluency, exam readiness or completion of a CEFR level.
Teach the writing system throughout
Introduce Hangul with meaningful speech from the beginning. Use romanization only as a temporary, optional aid and fade it. Add short reading and keyboard tasks; speech practice alone will not teach literacy.
Publish specific evidence of progress
Each weekly check includes a new role-play, a new listening recording, a reading item and recall of older material. Capture the first attempt before showing hints; then let the tutor teach and retry.
Japanese is a separate curriculum
Reuse exercise components, memory, scheduling and the world. Author a Japanese path around practical Can-do goals, kana alongside useful speech, gradual kanji, politeness and listening. Japan Foundation’s Irodori and Marugoto provide relevant curriculum references. [12] [13]
Commission Japanese review, reference audio and assessment examples separately. Do not translate Korean lessons or promise kana mastery in one week. Validate the English/Japanese voice experience before launch.
A focused route to learners.
The initial niche is Tamil-speaking adults who already enjoy Korean culture and are comfortable learning through English. Tamil creators and community connections give us an audience we can reach and understand.
Japanese and Korean ranked fourth and sixth globally in Duolingo’s 2025 report. This is a platform demand signal, not a count of our addressable customers. [7]
CJ ENM’s September 2025 Amazon MX Player deal announced 18 series with Hindi, Tamil and Telugu dubbing. Localized distribution supports the cultural-interest thesis. [8]
Netflix’s 2026 Made in Korea centres a woman from Tamil Nadu travelling to Korea. It is a useful cultural reference, not evidence of paid learning demand. [9]
Make the language contract explicit
Every ad and landing page says “Learn Korean in English.” Tamil can deliver the creator’s hook, testimonial or community message. Show the actual English coaching and Korean practice before signup.
Ask whether learners can follow a short English explanation during onboarding. Tamil identity or schooling does not establish English comfort. Record who the first version excludes so expansion decisions use evidence.
What we still have to learn
How many people want a course rather than occasional phrases; whether the character improves return; whether learning transfers to new situations; and what users will pay. The family conversation is a useful interview lead, not market validation.
English is the chosen delivery language for v1. We will test voice quality on our actual English–Korean tasks rather than making a broad claim about Tamil technology.
| Reference | What already exists | Implication for Morrow |
|---|---|---|
| Tolan | Everyday companion, memory and expressive embodiment. [1] [2] | Study continuity and personality; build explicit curriculum and assessment around our learning goal. |
| Speak | Korean and Japanese for English speakers; curriculum, spoken practice and AI conversation. [3] [4] | Direct competition. A learn–practise–apply loop is a baseline, not a unique claim. |
| Teuida | Human-written and filmed scenarios with speaking practice and pronunciation feedback. [5] | Our examples must sound natural and teach a useful role in a situation. |
| Duolingo / Lily | Japanese and Korean Video Calls, with beginner supports and feedback. [6] | A character plus voice chat already exists. Test the whole experience against learners’ current habits. |
The differentiation hypothesis: a persistent personal companion, evidence-based next steps and a world tied to practical situations can make English-led practice more compelling for this first audience. We must earn that preference through learning quality and retention.
A marketing test that shows the product
“You know the scene. Try the line.”
A Tamil creator introduces a familiar kind of café moment, then shows English coaching and an original Korean exchange in the app.
“My first order in Korean.”
Show an honest first attempt, a helpful correction and a second attempt. Keep the learner’s improvement central.
“It remembered what I struggled with.”
Show a real next-day callback and a targeted exercise. Demonstrate what the memory does for the learner.
Proposed 30-second creative: 0–5s Tamil hook → 5–12s introduce the companion → 12–23s show English explanation, Korean attempt and feedback → 23–30s invite viewers to the English-led Korean pilot. Creator copy and footage are commissioned after the product slice works.
The foundations are real. The tutor is the next build.
The local code audit shows useful reusable components. It also reveals the work required to turn a character demo into a reliable Korean learning product.
| Layer | Existing foundation | Required for the pilot |
|---|---|---|
| World and character | Three.js planet, rooms, navigation, reactions and narrated story opening. | Original character; a small coherent course route; lighter presentation for slower phones. |
| Voice activities | Cloudflare voice loop and English kitchen, picture and song activities. | Korean recognition, mixed English/Korean handling, turn control, retries and language-reviewed audio. |
| Tutoring | Lesson state machines and permissive word matching. | Objectives, accepted responses, feedback rules, hint ladder, skill records, spaced review and checkpoints. |
| Persistence | Session-local learner name and first-lesson completion in the tutor. | Durable learner account, cross-session progress, review queue and recoverable session state. |
| Companion memory | Mivuni summaries, profiles and nightly journals in another product domain. | Integrate relevant personal memory; keep it separate from skill evidence; add edit/delete controls. |
| Content and measurement | Story/audio production patterns and visual assets. | Reviewed Korean course, approved audio, content versions, event tracking, cohort reporting and teacher-rated assessment samples. |
Audit references: bt21-demo/voice-server/src/conversation.js currently uses flux-general-en; voice-server/src/lesson.js uses transcript word matching. tutor/main.js stores session completion. Mivuni summaries/journals are in mivuni-api/src/jobs/. These establish reusable code, not a completed Korean tutor.
The lesson engine stays in control
An authored lesson contains its objective, prerequisites, English explanation, target phrases, accepted variants, common errors, hint ladder, reference audio, exercise variants and rubric.
The sequence is learner attempt → recognition with uncertainty → task evaluation → feedback/retry → structured evidence → next activity. The model supplies bounded dialogue and proposed feedback. Only the lesson engine updates progress.
A reliable content pipeline
Teacher defines the outcome → draft lesson → qualified Korean review → audio review → check each answer/feedback path → small learner trial → publish a version. Review register and cultural context, not just literal translation.
Spend live generation where it helps
Pre-render lesson models, repeated instructions and story beats. Use short live dialogue for personalization and role-play, initially capped at three minutes per learner per day. Keep a guided fallback when a service fails.
Inworld documents Korean/Japanese TTS support; that makes it a candidate, not proof that any chosen voice is suitable. Cast and review the voice on the actual curriculum. Confirm storage and reuse rights for generated audio in the vendor agreement. [19]
Test mic permissions, noisy audio, interruption, network loss and small Android screens. Prefer push-to-talk and visible transcripts in the pilot. Give uncertain recognition a neutral retry path.
Keep the first version small
- One original guide and one reusable setting, with a small set of expressive reactions.
- Seven days of reviewed Korean at the first gate; grow to the four-week pilot when the teaching and speech work.
- English explanations, target-language practice, persistent progress, review and one clear weekly checkpoint.
- Public pilot for adults 18+, with visible memory controls and a documented audio/data retention policy.
- Defer a second language launch, an open-ended companion product, elaborate daily story films and native-store packaging until the learning loop is reliable.
Earn the next stage.
Dates are planning targets from a proposed 10 November start. Reviewer availability and voice quality are first-week dependencies. If they slip, reduce scope or move the gate before increasing acquisition spend.
- 10–16 Nov · prove the hard parts
One complete lesson, end to end
Agree the curriculum map and original character direction. Engage a Korean reviewer. Test English/Korean voice and recognition with target learners. Define events, skill records and assessment examples.
- 17–30 Nov · first release gate
A reviewed seven-day experience
Five core lessons plus review/checkpoint sessions, persistent progress, basic scheduling, one bounded role-play and one companion callback. Test on ordinary Android phones before inviting the first cohort.
- December · 20–50 external learners
Find teaching and reliability failures
Observe first sessions and next-day returns. Extend content toward 20 core lessons and four checkpoints. Use family testers for usability separately from the external outcome cohort.
- January · expand only after the gate
Measure the complete experience
Run small creator/paid tests into a clear English-led signup journey. Aim for 150–300 eligible learners if acquisition cost permits; this is a recruitment target, not a forecast. Run the companion comparison if sample size permits.
- February · read mature cohorts
Learning, return and willingness to pay
Read only day-30 windows that have fully closed by the decision time: enrollment must be at least 744 hours earlier (28 January or earlier for a 28 February gate, matched by time of day). Interview returners and dropouts. Test price interest with clearly labelled offers; actual conversion requires a later real paid test.
- 28 Feb · investment decision
Continue, revise or stop
Deliver the pilot, reviewed content, source handover, cost records and cohort evidence. Continue with one named improvement if evidence is promising; scope extra time separately. Japanese follows a successful Korean learning and retention gate.
Deliverable at the February gate: a usable web pilot and an honest evidence package. A Play Store release, 60 days of content and a second language are separate milestones, rather than assumptions inside the first build.
People return. They can do more.
The following are proposed decision thresholds to agree before recruitment. They are operating targets for our pilot, not industry benchmarks or proof of product–market fit.
| Signal | Proposed continue target | How we count it |
|---|---|---|
| First-session value | 60% complete the first teaching-and-retry sequence. | All eligible new users who start onboarding; report drop-off at each step. |
| Day-7 return | 30% return for a meaningful learning attempt. | All eligible assigned/onboarded users; at least one task attempt in the 24-hour window starting 168 hours after enrollment. |
| Day-30 return | 15% return for a meaningful learning attempt. | The analogous window starting 720 hours after enrollment, only for cohorts with a complete window. |
| Learning | At least 60% of assessed learners meet the pre-set task rubric at day 28, with improvement over baseline. | Use equivalent fresh tasks. Report everyone invited, assessment completion and missing outcomes; require at least 70% assessment coverage for a firm read. |
| Voice reliability | At least 95% of attempts complete without a service failure; fewer than 5% of sampled acceptable responses are falsely rejected. | Separate transport failures, recognition uncertainty and teacher-reviewed language judgments. Publish sample counts. |
| Cost and value | Variable usage stays within the operating envelope; returning users can explain what they value and react to a concrete price offer. | Record cost per learning session and per returning learner. Price-interest responses are not paid conversion. |
A learning check we can trust
Before the pilot, define parallel baseline/day-28 tasks and a rubric: accomplish the request, use appropriate taught language, and do it without sentence hints. Include listening and reading as well as speaking.
A qualified Korean teacher rates a blinded sample and all disputed cases. Report results for the original cohort alongside the assessed subset; users who disappear must remain visible in the outcome report.
The main progress claim is specific: “I can handle this new café exchange without the model.” Conversation length or lesson completion alone does not establish learning.
Test the companion package fairly
Randomly assign eligible learners before exposure. Both arms get the same course, voice quality, world rewards, reminders and access time. The comparison changes persistent personal callbacks and the journal; the control keeps a friendly lesson guide.
This estimates the effect of that companion package, not every Tolan-inspired element separately. Read retention together with learning and cost. Do not intentionally give the control a dull or broken world.
Illustrative power calculation: detecting day-7 return rising from 20% to 30% needs approximately 294 assigned users per arm at 80% power and a two-sided 5% test (normal approximation, rounded up). Our smaller initial cohort is directional. Set sample size using the observed baseline before the larger test; an inconclusive result is not “no effect.”
Decision rule: continue when return targets, learning evidence and reliability clear the agreed bars at sustainable measured cost. Pause acquisition to repair teaching or voice failures. If day-7 stays below 15% or day-30 below 7% after a meaningful fix and enough mature observations, reconsider the proposition. With small samples or wide uncertainty, record “inconclusive” and explicitly scope any further test.
A controlled pilot, with content review funded.
Keep the earlier $4,000 monthly proposal, with a revised working allocation that explicitly includes language review. Figures below are planning allowances, not vendor quotes or an approved commitment.
November–February: up to $16,000 across four monthly gates. Spend can stop at a gate; there is no assumption that every month must be used.
- Build and integration · Gokul
- $2,000
- Marketing experiments
- $1,000
- APIs, hosting and production tools
- $700
- Korean review allowance
- $300
The reviewer allowance must be checked against a real quote in week one. If review or original-character production costs exceed the envelope, reduce content/polish or agree a revised scope before committing spend. Japanese content and review are outside this pilot.
Release the marketing budget in stages
Working monthly split: $300 creator trials, $400 paid creative tests, $200 learner recruitment/incentives and $100 reserve. November emphasizes interviews and early hook tests; keep unneeded acquisition budget unspent until the product is ready.
Track creator/ad → landing visit → eligible signup → first lesson → day-7 learner. Optimize cost per learner who returns, rather than video views or raw installs. Use separate links for each creative and cohort.
College groups, K-culture communities and institutes are possible recruitment channels, not confirmed partners. Reuse real app captures; avoid assuming a costly brand film is necessary.
Make usage economics measurable
Variable cost = recognition minutes + generated speech + model input/output + optional assessment + storage/delivery. Cached audio still has production, review and delivery costs.
Illustration, not a quote: 300 learners × 12 active days × 3 live minutes = 10,800 live minutes/month. At an assumed all-in variable rate of $0.02–$0.06/min, that is $216–$648 before fixed tools and content production.
At 1,000 similarly active learners, the same assumptions give $720–$2,160. The $700 tools/API allowance is not unlimited scale. Meter each component, cap usage and expand only from measured costs. Inworld’s pricing is one input; voice turns and other services change the total. [20]
Agree the product, then earn the expansion.
“I want to build an English-led Korean tutor with a companion people care about. We’ll start by reaching Tamil-speaking learners, prove that they return and learn, then use the same platform for a separately reviewed Japanese course.”
- Approve the direction: English instruction, Tamil-focused marketing, Korean first, one original companion. Morrow is the working name pending brand clearance.
- Approve a gated pilot: proposed November–February envelope, the revised review allowance and the seven-day first release before a larger course.
- Agree the success rules: specific learning tasks, return windows, cost ceiling and what evidence would justify a paid test.
- Confirm delivery dependencies: language reviewer, voice acceptance, original-character scope and access to the reusable code and assets.
- Record ownership and operations: ownership of the new brand, character, course and code; rights to existing components; infrastructure fees, learner-data responsibilities and handover. These need a written agreement.
Expansion after evidence: test paid Korean access with the returning cohort, expand marketing to other English-comfortable communities, then author Japanese. A future BT21 collaboration would require a separate licensing discussion and business case; no license price or partnership is assumed here.
The references behind the plan.
Research refreshed on 7 November 2026. Official product pages establish advertised features; they do not prove learning effectiveness. Learning studies inform design principles. Proposed scope, prices, thresholds and forecasts are identified as our assumptions.
- Tolan · FAQ
Companion positioning and conversational memory. Basis for the relationship reference.
- Tolan · Designing characters
Expressive nonhuman embodiment and the move toward a peer-like character. Basis for presence and personality.
- Speak · Supported learning languages
English-speaker Korean and Japanese courses, documented July 2026. Direct competitive overlap.
- Speak · Product approach
Curriculum, spoken practice and AI conversation. Vendor description, not an independent efficacy comparison.
- Teuida · Human-authored scenarios
Human-written/filmed practice with AI pronunciation feedback. Reference for situation-based teaching.
- Duolingo · Product highlights
Korean/Japanese Lily calls and speaking supports. Availability and packaging may vary by account.
- Duolingo · 2025 Language Report
Japanese #4 and Korean #6 globally on its platform. Reporting period: October 2024–September 2025; not Tamil Nadu market sizing.
- CJ ENM · Amazon MX Player agreement
4 September 2025 announcement: 18 dramas with Hindi, Tamil and Telugu dubbing. A distribution signal.
- Netflix · Made in Korea announcement
24 February 2026 announcement of the Tamil Nadu–Seoul story and March 12 premiere.
- Online King Sejong Institute · Curriculum
Introductory Hangeul/basic-vocabulary course and English instruction availability. Sequencing reference.
- King Sejong Institute · Standard curriculum
Speaking, listening, reading and writing scope. Reference only; our course is independently authored and reviewed.
- Japan Foundation · Irodori
Practical Can-do outcomes for everyday communication. Reference for the later Japanese course.
- Japan Foundation · Marugoto teaching guidance
Communicative activities, language knowledge and ongoing script learning; temporary romanization support.
- Karpicke & Roediger · Retrieval, 2008
Experimental foreign-vocabulary evidence for repeated retrieval and delayed recall; not an AI-tutor fluency study.
- Cepeda et al. · Spacing, 2008
Spacing effects depend on retention interval. Supports distributed practice, not one universal review schedule.
- Lyster & Saito · Oral feedback, 2010
Classroom meta-analysis supporting oral corrective feedback, with stronger effects for prompts than recasts. AI transfer is a hypothesis.
- Microsoft · Speech language support
Lists ko-KR and ja-JP for pronunciation assessment. Language support does not establish reliability for our learners.
- Microsoft · Pronunciation assessment
Assessment model and feature limitations, including en-US-only prosody. Basis for separating recognition from pronunciation claims.
- Inworld · Model documentation
Multilingual TTS capabilities. Review the exact model, voice and language combination before production.
- Inworld · Pricing
Current vendor rate card. Our usage illustration deliberately uses assumed total rates rather than presenting TTS pricing as the entire cost.
Local evidence: existing brief, supplied Tolan film frames, BT21 prototype code/assets and Mivuni memory/journal code. Unsupported enrollment counts, review percentages, licensing prices and event-dependent forecasts from the earlier brief are excluded.