Wahori.
A cozy life-sim where English is how you survive. We teach by making you need the language, inside a world people actually want to be in.
1.5 billion
people
“learn” English.
Most apps teach them to translate apple and collect streaks. Completing a tree does not help you order coffee.
The gap between your last green checkmark and your first real conversation is where everyone falls.
The apps know this. They’ve added chatbots and speaking drills. None of it is survival — which is the only context in which adults actually acquire a language.
Language is a survival skill, not a school subject.
This is not a new idea. It has been published, peer-reviewed, and replicated for forty years. No consumer product actually applies it.
- Krashen1977
Language is acquired, not learned — through comprehensible input.
- Swain1985
Output plus corrective feedback is necessary for fluency.
- Schmidt1990
Adults need to consciously notice a pattern to internalize it.
- Long1985
Tasks drive acquisition — not rule-first instruction.
Denotation. The literal dictionary meaning of a word.
Pragmatic meaning. What a word actually does in context — connotation, idiom, register. Wahori tracks both, separately, per learner.
A sample conversation
“Hey, what can I get started for you?”
“Can I get a large coffee, no sugar?”
“room for cream?”
بيسأل إذا بدك تتركي مساحة بالكوب للحليب. جاوبي 'yes, please' أو 'no, just black'.
“No, just black. Thanks.”
Two unlocks in the last eighteen months.
NPCs that run on the device.
Gemma 4 edge holds a character across turns at acceptable latency. Two years ago: impossible. Now: every player gets a café full of distinct people whose consistency comes from data, not scripted trees.
L1-calibrated pronunciation feedback.
wav2vec2 fine-tuned on non-native English, plus Goodness-of-Pronunciation scoring, distinguishes intelligible from confusingat per-phoneme granularity. An LLM can’t do this — it can’t hear.
Places of articulation — the mouth positions we encode per learner, per source language, per phoneme.
It’s not the LLM. Anyone can call a model.
The moat is the pedagogy, encoded as data. Reproducing Wahori requires a linguist-engineer pair willing to spend months on mock-scenario calibration. That’s the asset.
- i
100+ calibrated mock scenarios
vocab, grammar, curveballs, register, phonetic demands — authored by a linguist, reviewed by linguists
- ii
L1 tolerance matrices
per-phoneme three-band models for Arabic and Japanese — pedagogical data, not model outputs
- iii
Curriculum rules engine
data-defined rules for when to introduce, surface, and re-expose — tunable without redeploys
- iv
Four-layer tracker schema
vocab × meaning, grammar patterns, phonetic profile, competency per vertical — feeding every prompt
Real numbers, not TAM fantasy.
Working Holiday visas
issued annually across EN-speaking countries — highly motivated learners with a deadline
Adults relocating
each year, globally, where English is the daily operating system — work, healthcare, housing
The app-graveyard cohort
tens of millions who bounced off flashcards — SRS-fatigued adults looking for something that works
MVP focus: Arabic and Japanese. Maximum linguistic distance from English. If the system works for them, it works for Spanish, French, Mandarin, and the rest.
Where we are. Where we’re going.
- Phase 0now
Architecture locked · spikes in flight
15-spike program running: edge LLM feasibility, phonetic ML, 100-scenario calibration corpus.
- Phase 1next
Vertical slice
One scenario end-to-end — every system represented in its minimal form. Proof the architecture holds.
- Phase 2
Content scale-up
The remaining eight scenarios authored on the slice's rails. Linguist contractors onboard.
- Phase 3
Closed alpha
Real AR and JP learners. Measuring retention + learning gains with instrumented play.
- Phase 4
Public beta · next raise
Open the doors. Waitlist converts. Raise for art, audio, and full-game production.
Zach
founder · engineer · linguist
Years learning languages. Formal linguistics study. Years shipping software. Owns both pedagogy and engineering until the team grows — which is rare, which is why Wahori works.
Arabic & English. Working in Japanese phonology. Building the thing I wish had existed when I first tried to survive in a language that wasn’t mine.
Joining as we go.
- 01Contract linguists for the 100-scenario calibration corpus and L1 tolerance matrices
- 02Native-speaker voice talent (Canadian, NYC, Australian English; AR and JP for helpers)
- 03Artists + animators for the 3D full-game post-MVP
Raising on a SAFE to get through closed alpha.
use of funds
- 40%Linguist contractors100-scenario calibration corpus + L1 tolerance matrices for AR and JP
- 30%Cloud & inferenceGemma-class edge, hosted Helper model, phonetic ML pipeline through closed alpha
- 20%Voice talentNative-speaker recordings for TTS consistency across NPCs and helpers
- 10%Founder runwaySix months to get through Phase 3 without burning out
Nobody acquired a language from a streak.
Adults acquire languages by needing them.
Wahori gives them the need.