The interview is powered by Claude Haiku (Anthropic), selected for response speed and cost per interaction. A custom system prompt encodes 20 barrier categories drawn from DBE/ACDBE personal narrative requirements under 49 CFR Part 26 and Part 23.

As each topic develops, the model detects when a section has enough detail to draft and emits a <<<SECTION_READY>>> marker. The client parses that marker to assemble the narrative panel in real time, one section at a time.

Voice input uses the browser's Web Speech API — microphone audio stays in the browser and is never transmitted. Voice output uses Google Cloud Neural2 TTS for natural, professional-grade narration.

Both the Anthropic and Google API keys are stored as Netlify environment variables and accessed only through a serverless proxy function. No API credentials are ever exposed to the browser.

On demand, the tool generates a Microsoft Word document client-side using docx.js. Output follows Personal Narrative formatting standards: Times New Roman body, structured section headings, statement of purpose, barrier narratives, conclusion, and signature block — ready for submission to the user's certifying Unified Certification Program.

Deployed via GitHub → Netlify continuous deployment. Pushes to main trigger an automatic build and production release.

Governance: API credentials are isolated server-side in Netlify environment variables — never transmitted to the browser. The document template is fixed in code, so the AI cannot alter structural formatting or omit required sections. Interview transcripts and generated narratives are processed in-memory and are not logged or retained by this application; third-party API calls to Anthropic and Google Cloud follow their respective data handling policies.