Show the pipeline
Every intermediate step is visible, such as the transcript before the reply. Nothing is faked or swapped behind the scenes.
Starter kits
Three small apps that prove the SDK works.
Each kit is a single screen or script that sends a real request through the SDK to N-ATLaS. The three windows below are the kits themselves, running live. Clone one, swap in your own content, and run npm install && npm start.
01
02
Label demo scale
Fixture datasets and shortcuts are marked "demo dataset" or "sample only" in the interface itself.
03
One shared look
Same layout skeleton across all three, so they read as one toolkit, not three unrelated demos.
Kit 01
Citizen Services
A single-page form for local-language Q&A over a small civic-info dataset, such as how to renew an ID. The language selector drives both prompt framing and language context.
- SDK calls normalizeText() → chat()
- Shows Language-aware prompting with injected context
- Does not Persist data, authenticate users or act as a real civic-info system
Backend offline
ResponseThe N-ATLaS answer appears here.
Question after normalizeText()
Kit 02
Education
A single-page tutor. The level selector shows the same chat() call reframed by prompt context, so you can see the SDK's flexibility without any new feature.
- SDK calls chat() · reportIssue()
- Shows Instructional framing by language and level
- Does not Track progress or store learner data
Backend offline
ExplanationThe N-ATLaS explanation appears here.
Your question, this answer and your correction are sent to OpenAtlas as a correction for N-ATLaS (reportIssue()). Nothing else is stored.
Kit 03
Customer Service
Two ways in for a customer. A voice note is transcribed, triaged and given a drafted reply, so the whole pipeline is legible on one screen: voice in, transcript, draft response. Or switch to Text chat for an ordinary back-and-forth conversation with a support assistant, for customers who'd rather type.
- SDK calls transcribe() → normalizeText() → chat() · reportIssue() · multi-turn chat() · speak() (stretch)
- Shows Language-routed ASR feeding the LLM; a multi-turn conversation grounded in one business's policies
- Does not Ship audio playback unless NAIC confirms the TTS pairing. Its absence is not a bug.
Backend offline
Voice note ready
TranscribedThe N-ATLaS ASR transcript appears here.
Drafted responseThe N-ATLaS LLM draft appears here.
Response audioRead aloud by a separate text-to-speech model from N-ATLaS's drafted reply, unchanged:
Your voice note and the corrected text are sent to OpenAtlas as a correction for N-ATLaS speech recognition (reportIssue()). Audio paired with a human-checked transcript is the data Nigerian-language speech recognition is shortest of.
Chat with the support assistant of a made-up shop, Ada & Sons Home Goods. Its replies come from the N-ATLaS LLM, using only the shop's sample policies.
Hello! How can we help you today?
Run one against the live endpoint.
Point the kit at the OpenAtlas gateway with one environment variable. No separate deployment needed.