// nuwright
● live · built; in partner user testing

relationalAI

Builds a support bot's system prompt and exports reviewed conversations as training data.

open pe-rai.vercel.app ↗
job
calibrate support chatbots
domain
behavioural health · platform
runs
hosted · API
class
platform
scale
215 example conversations across seven support profiles
since
2026

One call

  1. 01Set upOrganisation details and the bot's style on eight dimensions
  2. 02GenerateSystem prompt and knowledge package generated
  3. 03UploadTranscripts parsed, mapped and pseudonymised
  4. 04AnalyseA rubric flags safety, tone and structure findings
  5. 05ReviewA reviewer keeps or drops each finding and signs off
  6. 06ExportTraining data in the organisation's own voice

In · out

inputs
  • organisation details
  • the bot's style on eight dimensions
  • conversation transcripts
outputs
  • system prompt
  • knowledge package for retrieval
  • quality findings report
  • training data in standard formats
  • signed reviewer attestation

Where the person is

A reviewer keeps or drops every finding. Crisis conversations are never exported, and if a safety check fails, nothing is exported.

code decides● yes · the final say is deterministic

What is measured

Against: Synthetic and probe conversation sets; a clinician review is planned.

7 of 9 acute crises caught, with no false critical alerts on the synthetic test setmeasured
Known gap: 42 of 60 distress conversations that never name a crisis raise no flag, so a clinician review stays in the loopmeasured
Every generated system prompt includes crisis handling, checked automatically on every buildbuilt

With

Project EvidentThe Gates Foundation

Stack · links

ReactViteTailwind CSSVercel functionsAnthropic API
last updated 2026-09-28 · from source, not marketingclaims carry their own status