Public AI Governance Documentation
Vendor and Model Register
Public provider, model-use, data-flow, and review-trigger register.
This document is a public governance summary. It does not publish raw AI instructions, exact internal eval cases, source code, tenant data, incident details, database schema, or provider-console settings.
Last reviewed: July 6, 2026
This register describes HireProxy's AI and platform providers at a public, publishable level. It is not a substitute for private vendor due diligence, contracts, data processing addenda, or security questionnaires.
AI Providers
| Provider | Purpose | Data categories processed | Current risk controls | Review trigger |
|---|---|---|---|---|
| Anthropic | Career-agent responses, fit analysis, interview prep, answer evaluation, coaching, extraction, and structured generation | Candidate-provided profile data, stories, work history, job descriptions, recruiter questions, prep context, answer transcripts, and generated outputs | Scoped request design, output limits, reliability monitoring, no custom model training by HireProxy | Model upgrade, provider policy change, enterprise sale, material AI-instruction or data-flow change |
| OpenAI | Voice transcription and text-to-speech for Interview Mode | Practice audio submitted for transcription, transcript text for spoken coaching, generated coaching scripts | Authenticated endpoints, entitlement checks, rate limits, input size limits, no saved practice audio file in HireProxy records by default | Speech or transcription model change, voice retention change, enterprise sale |
Infrastructure and Application Providers
| Provider | Purpose | Data categories processed | Current risk controls | Review trigger |
|---|---|---|---|---|
| Supabase | PostgreSQL database, Auth, Storage | Account data, career data, resumes, interview prep, practice transcripts, usage events, settings, storage files | Supabase Auth, Row Level Security, server-side service-role access, storage path scoping | New table, new public data surface, RLS policy change, enterprise sale |
| Vercel | Hosting and serverless runtime | HTTP requests, deployment artifacts, runtime logs, environment variables | Server-side secrets, runtime boundaries, deployment rollback capability | Runtime change, sensitive surface change, enterprise sale |
| Vercel KV | Rate limiting and ephemeral counters | IP-derived or user-derived rate-limit keys and request counters | TTL-based expiry, scoped keys | Rate-limit architecture change |
| Resend | Transactional email and abuse-report delivery | Email addresses, transactional email metadata, abuse reports | Transactional-only use, limited report payloads | Email flow change, abuse intake change |
| LemonSqueezy | Merchant of record and subscription billing | Billing customer IDs, subscription status, payment processor records | HireProxy does not store full card numbers | Pricing, billing, or subscription architecture change |
| Sentry | Error monitoring and provider-failure observability | Application errors, stack traces, request metadata, application context tags | No raw user-facing error exposure, scoped error context, provider-failure metadata | Monitoring configuration change |
| LogRocket | Product diagnostics on non-sensitive marketing surfaces when enabled | Marketing-page interaction data, with text/body/input sanitization and blocked sensitive paths | Disabled on admin, candidate, start, upload, auth, interview, rehearsal, and sensitive API surfaces; IP capture disabled | Telemetry scope change |
| Cloudflare Turnstile | Bot prevention on public start flows | Bot-prevention challenge signals | Required before public-start uploads or AI preparation continue | Public start flow change |
Public Model-Use Principles
- HireProxy does not train or fine-tune foundation models.
- HireProxy centralizes model selection so model upgrades can be reviewed.
- HireProxy uses smaller or lower-cost models for narrower tasks when appropriate, and stronger models for synthesis-heavy tasks.
- HireProxy uses bounded requests, rate limits, input limits, and output limits to reduce unnecessary processing.
- Model or provider changes should trigger a grounding/confabulation eval run and a changelog entry.