The concrete origins of GamiWays
GamiWays was not born from an abstract idea. It is the generalization of two functional prototypes, tested with real users, which each revealed the same fundamental challenges: latency, experience personalization, immersion and voice quality.
Two experiences, one shared engine.
Parle à AVA! and Le Dilemme Plastique are two radically different experiences on the surface — one narrative and cinematic, the other pedagogical and scientific. Yet they share exactly the same technical challenges: STT→LLM→TTS pipeline, conversational agent orchestration, session memory management, perceived latency and French voice quality.
This convergence gave birth to GamiWays: a generic platform capable of supporting both the Storygami experience (interactive cinema) and the Edugami experience (guided learning), sharing the same orchestration engine.
The experience
Dive into a dystopian thriller with "Parle à Ava !" — an immersive conversational audiovisual experience accessible on desktop and mobile. You are catapulted into a mountain chalet where the family of Emma, Max, Ava (9 years old) and Léo (15 years old) tries to survive a virus that transforms women into men (the "protogyny").
"Parle à Ava !" is the interactive extension of the film "Where is Ava?", a dystopian thriller in production (public presentation planned for 2026). Explore the film's initial context, discover the family's stakes and secrets, and prepare to live a personalised conversational story, full of twists and surprises.
Through conversational AI and hyperrealistic synthetic videos, engage in unique dialogues with the characters of this world on the brink. Ask questions, express your hypotheses and explore complex themes in a world in peril. Each conversation reveals a fragment of a possible reality, with all its contradictions and complexity, to confront you with your own values.
The objective of this digital project: to make you think. Explore your reactions to the themes addressed and experiment with your relationship to your own certainties and moral principles through this original interactive experience.
Key metrics
User journey
Tools used
Technical challenges → documented solutions
ElevenLabs vs 15 alternatives — quality, latency, French naturalness. AVA uses ElevenLabs eleven_multilingual_v2.
Real-time avatar platforms — lip-sync, body language, latency. AVA's next step: photorealistic avatar of Max.
ELO rankings, TTFA, latency budgets — the metrics that define AVA's voice pipeline performance.
Context Engine v2, Memory System, Game Master — the GamiWays generalization of AVA's pipeline.
Le Dilemme Plastique
2 prototypes — 70h of development — ~380 CHF
The project
Dilemme Plastique is an innovative project combining documentary expertise, artificial intelligence and active pedagogy to raise awareness of environmental issues related to plastic pollution. It is designed to evolve into the generalisable educational tool Edugami.
The project was born from the meeting between Peter Charaf, a documentary filmmaker who spent a decade (since 2015) travelling the world twice to document the plastic problem in partnership with the Race for Water Foundation, and the Memoways team. This collaboration resulted in a unique proprietary database: field-verified images, videos and testimonials covering environmental pollution, health impacts, recycling challenges and emerging solutions.
After a first deterministic prototype presented at the Semaine des Médias in French-speaking Switzerland in February 2024, the team made a major technological pivot: abandoning the linear no-code approach in favour of a system 100% powered by generative AI. The student can communicate with Peter Charaf's avatar via voice or text, in an environment designed to maximise pedagogical engagement.
Long-term vision: Edugami is an educational tool generalisable to other themes. "We replace plastic with something else, another theme, history, another social issue." — Peter Charaf
Two approaches, one pedagogical challenge
Peter guides 12–18 year-old students through the analysis of an image of the Place des Nations in Geneva to discover 6 hidden clues. Sessions ≤5 min, 24+ simultaneous. Pipeline: Deepgram live → OpenAI Assistants API → ElevenLabs.
Split-screen interface: Peter chat (left) + media panel (right). Peter orchestrated via Flowise (28 nodes). 20–30 min sessions, no account. Includes videos, embedded articles, Postgres persistence, admin console, PostHog analytics.
Latency optimisations — measured values
Measured locally (CHANGELOG v1.2.0 Light + Flowise)
Measured at server boot — paid once for all students (Flowise CHANGELOG)
Instead of 3–5s — measured (Light CHANGELOG)
Overall gains (Phase 1 + 2: -40 to -55% total latency) are design estimates documented in CHANGELOG v1.2.0, not PostHog measurements. The AVA latency dashboard (admin) collects real per-session data but no aggregated values are published in the repo.
Tools used and development costs
Tools: Replit Agent (vibe coding + hosting), Claude Code, OpenAI GPT-4o Assistants API + Whisper, ElevenLabs TTS.
Technical challenges → documented solutions
ElevenLabs Scribe vs Deepgram vs Whisper — WER, latency, French accuracy. Dilemme uses Scribe (WER 2.11%).
ElevenLabs eleven_multilingual_v2 — 1507ms avg vs 2373ms OpenAI. Cache LRU strategy for classroom use.
Interactive V2V diagram — visualize the full pipeline latency budget as used in both Dilemme prototypes.
Flowise orchestration limits, LLM latency, STT accuracy on children's voices — gaps that GamiWays addresses.
The same challenges, one answer.
Despite radically different use contexts, both prototypes converged on the same unmet needs. This convergence is the technical and strategic justification for GamiWays.
| Challenge | AVA (Storygami) | Dilemme (Edugami) | GamiWays solution |
|---|---|---|---|
| Pipeline latency | GM+Max parallelization, fail-open validator | TTS LRU cache, pre-generated resume, Deepgram live | Runtime State + SSE, <2s latency budget per layer |
| Session memory | summarize-session (Facts/Topics/Promises, every 4 turns) | additional_instructions as source of truth | Memory System v2 — episodic + semantic + procedural |
| French voice quality | ElevenLabs stability 0.6, speed 0.92 — 'Clear & articulate' preset | ElevenLabs Scribe STT WER 2.11% — TTS eleven_multilingual_v2 | Documented Voice Pipeline — 17 TTS + 10 STT compared |
| Agent orchestration | Autonomous Game Master (trust, triggers, game over) | Flowise 28 nodes — complexity = bottleneck | Context Engine v2 — 7 dimensions, headless Game Master |
| Expressive avatar | Gumlet video + dynamic triggers (R&D) | Student avatar (thumbnail) — Peter without video avatar | Epic C.1 — Expressive Avatar Integration (Phase C) |
| Data sovereignty | Supabase Edge Functions — server-side keys | Replit + Neon Postgres — classroom data | LLM-agnostic architecture, Exoscale Swiss GPU cloud |
Discover the product vision, target architecture and Core Engine build status.