Supported public question
Classify → search reviewed records → answer → cite → show limitation
Success: The user gets a plain-language answer and can inspect its source.
ECCU STEMNisa build · student workspace
Learn · build · prove
The capstone product
Meet Nisa: useful information with a visible boundary.
Nisa is the National Insurance Smart Assistant. The team is building one explainable product that answers from reviewed public evidence, cites its sources and knows when the institution must take over.
Nisa informs. The NIS Board determines.
Four journeys
A trustworthy product handles more than the happy path
Personal case
Classify → stop before search → explain boundary → official handoff
Success: No personal information is requested, inferred, stored or decided.
Missing or conflicting evidence
Search → detect insufficiency → say what is unknown → official next step
Success: Nisa does not fill the gap with a plausible invention.
Dependency failure
Detect failure → preserve the boundary → safe fallback → trace the error
Success: The product remains honest and useful when a file, model or network fails.
The product layers
- Interface
- Shows the conversation, citations, limits and recovery states.
- Service
- Coordinates one request without hiding product rules in the page.
- Authority
- Routes personal, unsupported and unsafe requests before downstream work.
- Knowledge
- Admits only reviewed records with source, date and confirmation status.
- Retrieval
- Finds relevant evidence and preserves provenance.
- Response
- Builds a plain-language answer, citation, limitation and handoff.
- Quality
- Tests behaviour, records safe events and makes failures explainable.
Definition of Done
- The behaviour works in the integrated product, not only in an isolated file.
- Automated checks pass, including the relevant negative and failure cases.
- Another student reviews the change and the author can explain the diff.
- Public facts include a reviewed source and date context.
- Personal or unsupported requests fail safely before retrieval or generation.
- No secret, identifying data or raw personal question enters code or logs.
- The GitHub issue or pull request contains the evidence needed to verify the claim.
Build one product outcome
The labs coach the decision. GitHub records the implementation.