Generative audio is extraordinary at the forms it was trained heavily on — Atlanta trap, chart pop, lo-fi — and thin on everything else.
Log drum patterns, talking-drum polyrhythm, taarab-tinged strings, the specific pocket of a Lagos highlife brass line. The musical vocabulary that matters most on the continent is the vocabulary these tools handle worst.
A DAW, a laptop, a subscription in a currency you hold, and the English fluency to prompt well. The fastest-growing music scenes in the world are the least served by the tools reshaping music production.
A track nobody hears is a track nobody makes twice. Standalone generators hand you a file and a dead end — no listeners, no feedback, no reason to come back tomorrow.
A creator types "something for a Nairobi night drive." Gemini 2.5 Flash expands it into instrumentation, tempo, mix character and arrangement — the dense description Lyria actually responds to. The vocabulary problem, solved at the prompt.
Each carries a written Lyria prompt, BPM range, style tags and moods. Five are African forms — Afrobeats, Amapiano, Afro House, Bongo Flava, Gospel — written with the specificity the model needs rather than a one-word label that produces mush.
Tracks land in a ranked feed with follows, comments, gifts, remix lineage, daily challenges and battles. The output has somewhere to go and someone to hear it.
A full Next.js application — not a prototype. Auth, payments, streaming, lifecycle email, an admin surface and eight scheduled jobs, shipped and running.
Prompt-to-track generation, 23 genre templates, AI ideation chat, song blueprints, vocal takes, extend and flip, stem separation and an in-browser stem mixer, audio upload.
Ranked feed with time-decayed Heat Index, follows, comments, notifications, per-track share cards, an embeddable player, and vertical 1080×1920 video export for Reels and TikTok.
A daily themed challenge, head-to-head battles, cyphers where the community writes and votes on lyrics over a posted beat, leaderboards, streaks, and a "Most Built On" board.
Credit wallet, virtual gifts between creators, credit staking on tracks you believe in, referrals, and a creator analytics dashboard.
Every generation on Kavosuri runs on Lyria. Every prompt that reaches Lyria was written by Gemini. Provenance is checked with SynthID before the track is published.
Non-Google models handle vocal synthesis and stem separation where Google does not yet cover the surface. No exclusivity is required by the Lab, and we would rather be precise about the seams: the generation core, the prompt intelligence and the provenance layer are all Google.
Production figures, 29 July 2026. Organic growth from a standing start in June, with no paid acquisition.
Final column is a partial week (2 of 7 days at time of writing).
35 of 102 tracks are built on top of another track, across 13 distinct parents. Collaboration is the behaviour the whole product is designed to produce, and it emerged on its own.
We audited the funnel against the production database instead of guessing. Credits weren't the constraint — 40 users were sitting on 50+ unspent. Three feedback loops were severed: nothing came back to a creator after posting, the ranking algorithm was mathematically dead so new tracks were buried at zero, and the retention emails hung off a webhook that never fired.
New Heat Index with sane decay, play-counting moved into the audio layer so it actually counts, follow and gift and milestone notifications, a daily challenge, streaks, and seeded follows at onboarding so the feed isn't empty on day one. Notifications have gone from 17 lifetime to 41; 7 users are on active streaks.
1 — IFPI Global Music Report 2026 (2025 figures). 2 — Music Ally / Africa streaming forecast.
Free daily allowance, subscription tiers, and one-off credit packs. Generation, extends, stem unlocks and gifts all draw from the same wallet. PayPal is integrated and working.
Gifts between creators, stem unlocks, and staking credits on a track you believe in. Small amounts, high frequency, creator-to-creator — the pattern that already works across African mobile commerce.
Charging for a product people don't return to converts a retention problem into a refund problem. Payments are wired and dormant. Current tier pricing is US-benchmarked and almost certainly wrong for the market — this is the question we most want the Lab's partners on.
Full-stack engineer, Nairobi. Building software since 2022; graduated November 2025. Specialisms: full-stack web development, AI model integration and API design — which is, more or less, the entire surface area of this product. Full-time on Kavosuri.
Next.js application and design system, Postgres schema and queries, Vertex AI and Gemini integration, credit ledger and payments, encrypted audio streaming, transactional and lifecycle email, admin tooling, and eight scheduled jobs. No agency, no contractors, no co-founder.
When retention looked wrong, the response was a full funnel audit against the production database — which found that the obvious answer (people are running out of credits) was false, and the real causes were three broken feedback loops. Then those got rewritten and shipped inside a week.
The genre taxonomy was written here, not licensed from anywhere. Bongo Flava is described to the model as "bright interlocking electric guitar lines, marimba and kalimba accents, taarab-tinged strings, warm East African coastal groove" — the level of detail that only comes from knowing the music. The first 73 users are creators in this market, and what they do with the product is the fastest feedback loop available.
Kavosuri went from nothing to a live product with real users inside eight months of graduating — auth, payments, streaming, admin tooling and a social graph, all shipped and running in production.
We already run the preview endpoint with a fallback. Dependable access to longer-form, higher-fidelity vocal generation is the difference between "make a 30-second loop" and "finish a song" — between a toy and a tool.
We shipped vertical video export last week because sharing is how a track finds an audience. It's currently a waveform over cover art. Veo turns every track into a music video, and every music video into reach.
What does it take to make a frontier music model render amapiano, bongo flava or highlife correctly? We have the creators, the genre taxonomy and the appetite to evaluate. We lack the model-side access to do it properly. The answer is useful to Google, not only to us.
Payments are built and dormant. The open question is what the right unit and price point are for creators across the continent — precisely the question the Lab's partner VCs have answered before.