Back to Projects
A
Engineering

Anthrovian

Full-Stack Developer

Next.jsTypeScriptPrismaPostgreSQLElevenLabsPWA

About the Project

Anthrovian is a platform for interactive African folklore — myths played as branching adventures rather than read as static stories. The flagship title, 'Sundiata's Rise: Awaken the Lion Within', adapts the Epic of Sundiata across 11 scenes and 6 endings, with choices scored against three philosophical variables from Mande thought: Badenya, Fadenya, and Nyama. It runs on a 'Living Canon' model — stories are co-authored with African storytellers and cultural custodians, with AI used to scale and branch the narrative rather than to author it. Built mobile-first as a PWA with ElevenLabs voice acting, ambient soundscapes, and a WhatsApp play surface alongside the web app.

What It's Made Of

Narrative engine

11 scenes · 6 endings · 3 Mande variables

The branching core: choices are scored against three philosophical variables from Mande thought — Badenya, Fadenya and Nyama — rather than a single linear score, so an ending is reached by the shape of a player's values across the story and not by a points total.

Web PWA

Next.js · mobile-first · installable

The primary play surface, built mobile-first and installable. Anonymous-first by design: a device token in local storage starts a playthrough with no signup at all, with email added later as an optional cross-device resume key rather than a gate on the front door.

Save & state layer

Prisma · PostgreSQL · schema-agnostic JSON

Full game state stored as opaque JSON rather than a normalised schema. Narrative content changes constantly in production — scenes rewritten, choices renumbered — and a normalised save schema means every edit risks invalidating playthroughs already in progress. Queryability was traded for saves that survive story edits.

Audio & voice

ElevenLabs · ambient soundscapes

Voice acting through ElevenLabs synthesis plus ambient soundscapes per scene, and a shareable ending card generated for social distribution so a completed playthrough has something to post.

WhatsApp play surface

Own session model

The story playable inside WhatsApp with its own session handling, meeting the target audience on the channel they already use rather than asking them to install something. Same narrative, different interaction model, which is why it needed a separate session layer rather than a wrapper around the web client.

Living Canon pipeline

Human-authored canon + AI branching

Stories are co-authored with African storytellers and cultural custodians, and AI is used to scale and branch the narrative, never to author it. That boundary is structural in the pipeline: human-written canon and generated variation are kept clearly separated rather than blended into one content pool.

Key Highlights

  • Designed the branching narrative engine: 11 scenes, 6 endings, scored against three Mande philosophical variables rather than a single linear score
  • Built anonymous-first play — a device token from localStorage starts a playthrough with no signup, with email as an optional cross-device resume key added later
  • Stored full game state as schema-agnostic JSON so saves survive story revisions and save-format bumps instead of breaking mid-playthrough
  • Integrated ElevenLabs voice synthesis and ambient soundscapes, with a shareable ending card generated for social distribution
  • Extended play to WhatsApp with its own session model, meeting the target audience on the channel they already use

Technical Challenges

The save format was the decision that mattered. Narrative content changes constantly during production — scenes get rewritten, choices renumbered — and a normalized save schema means every one of those edits risks invalidating playthroughs already in progress. Storing the whole game state as opaque JSON gave up queryability in exchange for saves that survive story edits, which is the right trade when the story is the thing still moving. The other constraint was cultural rather than technical: AI branches the narrative but never authors it, so the pipeline had to keep human-written canon and generated variation clearly separated.