Imran M N
Resume
Back to Blog

Why Lingosee ships without an LLM

4 September 2026 · 1 min read

⇄ Paired with No. 007: LingoSee - Language Learning

Most "AI language app" pitches are really a chatbot with furigana. Lingosee started as one too: an LLM-driven NPC engine with hidden intent, a model playing every character. What's live at lingosee.imranmn.life is a different, smaller thing: a fully scripted, deterministic roguelike with zero backend calls and zero LLM in the loop.

WHAT'S DONE

Six real scenarios stand between you and functioning independently in Japan: immigration at the airport, the ticket machine, platform announcements, the bus stop, a convenience-store checkout under a timer, and a real-estate negotiation. Every NPC is a content object, not a prompt: dialogue, branching, and scoring are authored as data and resolved through dedicated NPC, scoring, and Sensei modules, all covered by an automated acceptance-test suite. Voice runs both ways through the browser's own Web Speech API, with typing as the always-available fallback. Three strikes end a run; a meiwaku (social-debt) meter tracks how much you're getting away with; a spaced-repetition deck holds the vocabulary between sessions; and a Sensei character reviews your full transcript after each scenario to explain what actually happened.

WHY SCRIPTED OVER LLM-DRIVEN

A scripted engine trades away some flexibility for something a game with a testing claim actually needs: every rule the game enforces, like rewarding repair over lucky guesses, is deterministic content you can write acceptance tests against, not a prompt you're hoping holds up.

WHAT'S NEXT

More scenarios past the first six, and a harder look at whether an LLM belongs anywhere in the loop now that the scripted core is proven out, as an option layered on top, not a replacement for it.

Imran M N is a product & strategy consultant (PwC, ex-Deloitte, IIM-B), working across cloud economics, product discovery, and applied AI. Get in touch →

Resume ↓Email