We believe people should not have to leave their language behind to access knowledge, opportunity, intelligence, or the tools that shape tomorrow.
Access should not require linguistic surrender. No one should be locked out of the world's knowledge because the most powerful technologies do not speak their language. Everyone should be able to learn, create, solve problems, and help shape the future — regardless of language or location.
Build Bantu-language AI infrastructure at family scale. We systematically curate Bantu family data and build the infrastructure, evaluation systems, linguistic resources, and model-improvement tools that help AI systems master Bantu languages with the competence they show in English.
No Bantu language left behind. No LLM left behind. We envision a world where every community can access, use, and contribute to the world's knowledge on its own terms — with fluent, competent, culturally grounded AI available in the languages people actually speak.
An FSI is the operating alphabet of a language — the complete, closed, finite set of its legal syllables. It is infrastructure, not a dataset: like the 26 letters, you do not size it by volume. A Bantu child learns syllable-by-syllable (e.g. ba · be · bi · bo · bu); those syllables form the Full Syllable Inventory — the closed set from which every native word is built. Standard AI models, trained on flat text, cannot recover or self-certify this system and confidently hallucinate phonetically impossible structures — and no frontier model can produce the operating alphabet of even one of the 500+ Bantu languages unaided.
Standard BPE tokenizers split agglutinative Bantu text along arbitrary frequencies, tearing prefixes from stems. Injecting FSIs as fixed vocabulary tokens cures this fracturing, reducing tokens-per-word by up to 40% and lowering semantic perplexity.
The syllable row ca ce ci co cu is written identically but pronounced completely differently: a soft consonant in Bemba, dental clicks in Zulu and Xhosa, and not used at all in Swahili. An FSI maps these spelling-to-sound boundaries deterministically.
Verify the FSI gap on your own stack. We run a 75-day pilot across 3 languages you choose, measured against pre-registered, contract-locked bounds: ≥80% recall, ≤5% fabrication, and ≥95% native linguist audit correctness.
FSI = NSI ∪ ASI
NSI (Native Syllable Inventory): Closed set conforming to the language's own phonological template: \(\sigma o (N)(C)(H)(G)V\), where only the vowel \(V\) is required (nucleus) and there is no coda.
ASI (Augmented Syllable Inventory): Mapped syllable set capturing loanwords, modern technical terminology, and code-switched adaptations.
459 of the 500+ Bantu languages now have their operating alphabet defined — each one complete and foundational on its own (coverage of the foundational layer, not a measure of how much data exists). This is the layer every Bantu word is spelled and spoken on. Loading an FSI also cures tokenizer fracturing — but that efficiency is a consequence; the point is the primitive now exists, ground-truth verified, where no model could produce it. Verify the gap on the public Alphabet Test.
Every one of those FSIs, assembled onto one canvas. The atomic unit becomes the syllable cell — an onset × a vowel — and each cell registers how much of the family shares it: a periodic table of Bantu syllables, grounded in consented native audio. The dense core is the shared Bantu backbone; the sparse edges are the onsets that fingerprint individual languages.
Public preview shows family-level coverage. Evaluators get the full per-language Map; Founding Partners get the audio chorus, comparative lens, and export.
The Economics of the Bundle
A single consented Bantu–English code-switch recording is not one labelled example. Captured in one continuous take, its switch point is logged as switch_ms — so the record self-segments with no manual annotation, and one take becomes the ground truth for six things frontier labs otherwise buy from six separate vendors.
Each product is a standard + data layer an AI lab consumes — the FSIs and BTS standards (infrastructure), and the consented audio corpora (Tone, Numbers, Health). One credential works across all of them, and every code-switch recording feeds six of them (above).
Prove the gap free, verify the data on an allow-listed set, measure it on your own held-out data in a paid pilot, then license the whole ecosystem. One credential carries across every product at every level.
See exactly what each level unlocks →
75-day pilot across 3 languages of your choice. Evaluates FSI token integration and performance, with 100% of the cost credited towards your first-year subscription.
Request Validation Spec →The whole living program — every domain and language (syllables, verbs, tone, health, numbers, and each one added next), the full consented audio corpus, all exports and APIs, with production rights. Not a dataset: a subscription to a platform that keeps growing. The $1.75M/yr is the founding rate, locked two years — it rises afterward.
Initiate Contract →Access is invite-only for commercial partners and researchers. Once logged in, your browser session grants access to all ecosystem subdomains automatically.