Izwa

Learned-data infrastructure and applied AI.

Izwa turns raw observations into verified datasets, specialised models, and deployable intelligence. African languages—beginning with siSwati, isiZulu, isiXhosa, and isiNdebele in the Nguni family—are the first research and commercial focus.

We are not trying to build the biggest model. We are making more of the world learnable.

What the system does

collect and preserve
  → validate and route
  → create machine drafts
  → review and adjudicate with people
  → publish versioned datasets
  → train or adapt models
  → evaluate independently
  → deploy demonstrated capabilities
  → turn failures into the next data priorities

The logical record connecting source evidence, permissions, annotations, review history, representations, model behaviour, and releases is the Learned Data Unit (LDU).

Product status

Area	Current status
Collection and human review	Working foundation
Versioned datasets and evaluation cases	Working foundation; no public dataset release yet
Dataset licensing and evaluation services	Commercial foundation; demand still to be validated
LoRA adapter licensing	Proposed; requires a reproducible improvement result
Custom model adaptation	Proposed
Private deployment	Proposed
Hosted model API	Proposed

“Foundation” means supporting workflow code exists. It does not mean a sellable product or model capability has been demonstrated.

Repository map

Path	Responsibility
apps/api	Canonical Hono API Worker and same-origin web delivery
apps/web	Public site, contribution studio, product and contributor pages
apps/annotator	Human correction, review, and adjudication surface
apps/mobile	Voice-first and field contribution surface
packages/domain/src/izwa	LDU, quality, provenance, and refinery domain logic
packages/datasets	Dataset and release contracts
packages/annotation	Human task and review contracts
packages/evaluation	Protected evaluations and benchmark logic
packages/embeddings	Representation lineage and retrieval support
packages/taxonomy	Language, variety, task, and phenomenon taxonomy
packages/db	D1 schemas and migrations
workers	Background ingestion, refinery, session, and workflow workers
legacy/sandbox	Archived template worker; do not extend

Start locally

Requirements: Bun, a supported Node.js runtime, and Wrangler authentication for Cloudflare-backed operations.

bun install
bun run typecheck
bun run test
bun run dev

Useful focused commands:

bun run dev:api
bun run dev:web
bun run build
bunx wrangler dev --cwd apps/api

Local services use the bindings declared in apps/api/wrangler.jsonc. D1 migrations live in packages/db/src/migrations.

API

The full-stack deployment serves the SPA and Hono API from one origin.

Endpoint	Purpose
GET /api/izwa	Discover the API and current Nguni focus
GET /api/izwa/direction	Canonical direction and operating stages
GET /api/izwa/capabilities	Foundation and proposed offering status
GET /api/izwa/languages	Current Nguni focus languages
GET /api/izwa/languages?scope=research	Wider research taxonomy
GET /api/izwa/contract	Active language quality contract
GET /api/openapi.json	OpenAPI 3.1 document
GET /redoc	Browser API reference

Live deployment: izwa-api.mkfixpro.workers.dev

Deploy

The canonical full-stack target is izwa-api. It packages the current apps/web/dist assets and runs API routes first for /api/* and /health.

bun run typecheck
bun run build
bunx wrangler deploy --cwd apps/api

Apply remote D1 migrations separately when a release contains migration files:

bunx wrangler d1 migrations apply izwa-db --remote --cwd apps/api

Do not deploy legacy/sandbox as part of Izwa.

Documentation

Start with the documentation index.

• Canonical company direction
• Product guide
• Product and commerce
• System map
• Developer guide
• Evaluation programme
• Deployment guide

Release standard

Izwa does not claim a capability because a route, model call, or demo exists. A product claim requires permitted data, a versioned dataset, a protected evaluation, reproducible configuration, comparison with an unchanged baseline, and results reported for the specific language and task.

License

MIT. Dataset, model, source-material, and contributor rights are governed by their own explicit terms and are not granted by the repository licence.