The Hybrid Sandwich.
Rust where latency matters. Python where reasoning matters.
The Narvo Intelligence pipeline is engineered against a single structural pressure: deliver verifiable African-language audio at sub-three-second synthesis latency, without the cost profile that has killed every previous attempt.
Rust Gateway (Axum)
Validates auth tokens at the perimeter, enforces per-key sliding-window rate limits, and terminates TLS against Cloudflare. Intercepts 100 % of inbound traffic; Python never sees unauthenticated requests.
Python / FastAPI
Safe-harboured behind the Rust gateway. Runs narrative synthesis (Groq → Gemini Flash → Workers AI fallback), RAG-based provenance verification against Africa Check + Dubawa, translation, and LangGraph-orchestrated multi-step reasoning.
Rust Delivery Layer (Axum)
Encodes audio into dual output — MP3 for persistent storage and WhatsApp-shareable clips, Opus for streaming. Uploads to Backblaze B2 through Cloudflare Bandwidth Alliance — structurally zero egress fees across terabytes of media.
TTS provider hierarchy
- Primary core: YarnGPT — Nigerian English, Yorùbá, Hausa, Igbo, Pidgin. Locked for the 5 launch languages.
- Expansion challenger: OmniVoice — Zero-shot Swahili, Arabic, and Francophone deployments pending 82+ benchmark pass.
- Redundant fallback: OpenAI TTS · Google Cloud TTS · Kyutai Pocket TTS — English-only latency staging; never used for first-tier African languages.
A new voice enters the production roster only after passing a blind 3-panel review (general listener, journalist, linguist) at a combined 82+ prosody score. Scores under 80, or any mispronunciation of a canonical Nigerian proper noun (Abuja, Lagos, Naira), are hard blockers.
Storage & delivery economics
Audio bandwidth is the biggest cost risk in this category. Narvo dodges it structurally: assets are pushed exclusively to Backblaze B2 and served through the Cloudflare CDN edge, exploiting the Bandwidth Alliance for effectively zero transit fees. The architecture is not an optimisation — it is a prerequisite for unit economics at ₦50/day.
Verifiable proof points
- 5 Nigerian languages at MVP launch — YarnGPT benchmarked on a 90-point prosody rubric.
- <3s synthesis P95 latency — Groq LPU inference + async TTS pipeline.
- Zero CDN egress on audio — Backblaze B2 + Cloudflare Bandwidth Alliance.
- Provenance score per story — pgvector cosine search against Africa Check + Dubawa.
- Multi-tenant B2B ready — Per-API-key isolation + quota from Phase 0.
- Data-safe offline listening — IndexedDB-cached MP3; no cellular auto-download.
Related reading
For the full shared-infrastructure specification (voice benchmarking rubric, RAG indexing schedule, language rollout gates), see the parent engineering documentation maintained internally. A redacted public mirror is linked from the press page on request.