For a merchant managing 30 transactions a day — in a loud Dakar market,
sometimes with a child on their back — the keyboard isn't a friction point.
It's a barrier. Kuma removes it entirely.
We use AI to translate voice into structured data, in the languages your users
actually speak. No keyboard required.
60M+
micro-businesses in Africa's informal economy running on paper
$400B
unmet financing need for informal traders across the continent
<3s
voice to confirmed transaction — our V1 target latency
3
languages supported — French, English, Wolof — with code-switching
How it works
Voice in. Structured data out.
Six steps, under 3 seconds. Every word spoken becomes a validated, structured
transaction — no keyboard, no menu navigation, no friction.
01
Push to talk
One tap. The merchant speaks naturally in Wolof, French, or a mix — exactly how they talk in the market.
audio capture → base64
02
Speech recognition
Google STT v2 with wo-SN support transcribes Wolof and French simultaneously — code-switching handled natively.
wo-SN + fr-SN multilingual
03
Wolof parsing
Our custom Wolof number parser (based on Guérin 2021) extracts amounts: "ñaar junni" → 2,000 XOF instantly.
parse_amount("ñaar junni") → 2000
04
Intent classification
Hybrid search (RAG + keyword) against our CommerceIntents catalog in Spanner — semantic understanding, not keyword matching.
RECORD_CREDIT · confidence: 0.94
05
Business validation
Rules run before any write. Amount > 0? Client identified? Permission valid? The LLM never executes directly.
validate() → { valid: true }
06
Merchant confirms
The confirmation screen shows what was understood. The merchant validates — then Kuma writes to your API. Always human-in-the-loop.
→ your_app.record_transaction()
The technology
We built the Wolof number parser.
Based on peer-reviewed linguistics research (Guérin 2021), our parser covers
all Wolof numbers from 1 to 1 billion — including the commerce-specific
dërëm unit (1 dërëm = 5 CFA francs).
benn
1
One · determinative
ñaar junni
2,000
Two thousand
2 000 XOF
juróom fukk
50
Five × ten
50 XOF
fukk ak juróom ñaar
17
Ten + five + two
17 XOF
fanweer
30
"Jour-lune" · lunar month
30 XOF
téeméer
100
One hundred
100 XOF
ñaar téeméer dërëm
1,000
200 × 5 CFA dërëm
1 000 XOF
juróom-benn-i junni
6,000
Genitive construction
6 000 XOF
Why this matters: In Dakar markets, prices are often quoted in Wolof even by French-speaking merchants.
A merchant saying "ñaar junni" means 2,000 XOF. Generic ASR returns garbled text.
Kuma returns { amount: 2000, currency: "XOF" }.
Technical stack
Built on Google Cloud, designed for Africa.
Infrastructure that scales from 10,000 to 10 million merchants without a refonte.
Voice & AI
🎙
Google Cloud STT v2
wo-SN + fr-SN multilingual, code-switching
core
🔍
Hybrid Search (RAG + keyword)
Spanner embeddings + RRF scoring
core
🤖
Gemini 2.5 Flash
Strict JSON output only via ML.PREDICT
core
📖
Wolof Parser
Based on Guérin (2021) — 1 to 1B
core
Mobile
📱
React Native + TypeScript
iOS + Android, low-end Android optimized
ProBoutik
🔌
Offline queue (MMKV)
Transactions en attente si faible connectivité
ProBoutik
Backend & Data
⚡
FastAPI + Python 3.11
Async, Pydantic validation, OpenAPI docs
core
🗄
Google Cloud Spanner
ACID distributed DB + vector embeddings
core
☁
Cloud Run (serverless)
Auto-scale, zero infra, pay-per-use
core
📊
VoiceLogs → feedback loop
Every API call feeds parser improvement
moat
Google ADK
🔗
Sequential Agent Pipeline
Audio → STT → Parse → Classify → Validate
core
Who plugs in
One API. Any app. Any market.
Kuma is infrastructure — bring your app, we bring the voice layer. Your users speak, Kuma understands.
🏦
Fintech apps
Voice-enable merchant transaction recording for mobile money platforms and merchant banking apps across West Africa.
First deployment: ProBoutik × Orange
🌱
Agri platforms
Field workers record harvest, expenses, and inventory by speaking in their local language — no literacy requirement.
Phase 3 partner
🤝
NGOs & field ops
Data collection without keyboard barriers. Surveys, health records, census data — spoken in Wolof, stored as structured JSON.
Phase 3 partner
🤖
AI platforms
Add West African language support to existing voice agents on Retell, VAPI, or custom pipelines. Kuma as middleware.
API partner
🏛
Governments
Census collection, subsidy registration, and public services access for populations where typing is a barrier.
Phase 3
✨
Your use case
Building for West Africa or any emerging market where the keyboard was never the right interface? Let's talk.
→ Apply for early access
Roadmap
Three phases. One compounding moat.
Each phase makes the next one stronger. ProBoutik generates the data. The data makes the API better. The API opens every door.
PHASE 01 — NOW
ProBoutik × Orange Sénégal
Build the Kuma voice layer inside Digicaisse V2 — a new voice-centric merchant app distributed to 100,000+ merchants across Orange West Africa. Revenue sharing model. Sentreso owns the voice IP.
In progress
PHASE 02 — DEPLOYMENT
Data flywheel
Every ProBoutik transaction generates a gold-standard Wolof commerce training pair. Real merchant voices, real code-switching, real market noise. This dataset is Kuma's competitive moat — nobody can replicate it without the same deployment.
6 months post-launch
PHASE 03 — PLATFORM
Open API infrastructure
API pricing per call + revenue share on embedded deployments. Any fintech, NGO, agri platform, or AI service wanting West African voice support plugs into Kuma. ProBoutik is the proof point that opens every door.
6-12 months post-launch
Get started
Ready to add African voice to your product?
We're onboarding our first API partners now. Apply for early access and get free integration support for 90 days.