Public AI Governance Documentation

AI System Cards

Public system cards for the main HireProxy AI-enabled surfaces.

Last reviewed: July 9, 2026

These system cards describe the main AI-enabled surfaces in HireProxy at a public, publishable level. They intentionally omit source code, raw prompts, secrets, and tenant data.

1. Public AI Career Agent and Fit Analysis

Purpose: Help recruiters and hiring managers ask questions about a candidate's professional background, and help candidates understand role fit against a job description.

Primary users:

• Candidates who own and maintain their career data.
• Recruiters and hiring managers visiting a published candidate site.

Inputs:

• Candidate-provided profile, work history, stories, gaps, rules, uploaded resume text, contact links, branding preferences, and optional job description text.

Outputs:

• Career-agent answers, contact suggestions, job-fit summaries, requirement narratives, and evidence-based fit notes.

Human oversight:

• Candidates control what they publish and can edit or delete their own data.
• Visitors are told that AI can make mistakes and should confirm final details with the candidate.
• Report-abuse intake is available for misleading or false content.

Primary risks:

• Confabulation or overstatement of career facts.
• Employer/client confusion.
• Privacy exposure from candidate-provided material.
• Attempts to override intended AI behavior or extract system behavior.
• Misuse as an automated hiring decision tool.

Current controls:

• Grounding in candidate-provided career data.
• Evidence-oriented AI instructions and uncertainty fallback.
• Candidate instructions plus non-negotiable accuracy guardrails.
• Public terms state that HireProxy is not an automated employment selection, ranking, or rejection system.
• Per-IP, per-user, and per-subdomain rate limits where relevant.
• Tenant-scoped access controls and Row Level Security.

Known limitations:

• HireProxy does not independently verify candidate-provided career claims.
• AI responses can be incomplete or imprecise.
• Public candidate sites may include information the candidate chooses to make public.

2. Interview Prep Generation

Purpose: Convert a job description or target role into an interview preparation guide, study cards, likely questions, watch-outs, and evidence prompts.

Primary users:

• Candidates preparing for recruiter screens, hiring-manager interviews, peer interviews, panel rounds, technical rounds, or executive rounds.

Inputs:

• Candidate profile and career stories.
• Job description, target role, company name, interview stage, interviewer context, and optional company research.

Outputs:

• Prep guide, study cards, likely questions, story cues, role-specific gaps, watch-outs, and logistics guidance.

Human oversight:

• Candidate reviews the generated prep before using it.
• Candidate can edit interview context and regenerate prep.

Primary risks:

• Overstating role fit.
• Incorrect company context.
• Generic guidance that does not reflect the candidate's real evidence.
• Confusing adjacent experience with direct experience.

Current controls:

• The job description is treated as the source of truth for role requirements.
• High-risk capability claims require explicit candidate evidence.
• Unsupported company facts are omitted rather than guessed.
• Generated prep remains private to the candidate.

Known limitations:

• Company context from public sources may be incomplete or stale.
• The guide is preparation support, not career, legal, employment, or hiring advice.

3. Voice-Led Interview Practice and Evaluation

Purpose: Let candidates practice answers out loud, receive spoken coaching, retry stronger answers, and identify private career-story improvement opportunities.

Primary users:

• Candidates practicing interviews on desktop or mobile.

Inputs:

• Interview prep context, question text, candidate answer transcript, optional speech metrics, retry history, and the candidate's saved prep.

Outputs:

• Quick spoken feedback, full scored evaluation, follow-up questions, delivery metrics, retry focus, practice-ready next-pass guidance, and private story opportunity suggestions.

Human oversight:

• The user decides whether to retry, type instead of speak, refine a private answer into a public career story, or delete practice attempts.
• Raw practice answers are private by default and are not recruiter-facing unless the user chooses to refine and publish them.

Primary risks:

• Confabulated feedback or unsupported claims.
• Unfairly harsh or misleading coaching.
• Voice/transcription privacy exposure.
• Misuse as undisclosed live-interview assistance.
• Mobile hands-free interruption or session failure.

Current controls:

• Evaluation instructions require the AI to evaluate only the transcript and not invent facts.
• Full written feedback scores direct question relevance, answer clarity, candidate-owned action, result or impact, and story fit.
• Quick spoken feedback is concise, scoreless, and separate from full rubric feedback so speed gains do not replace the detailed evaluation.
• Terms prohibit undisclosed live-interview assistance.
• Voice features require authenticated access, entitlement checks, rate limits, and session ownership checks.
• When free voice practice ends, the product presents an upgrade path while preserving typed rehearsal fallback.
• Practice audio is processed for transcription and speech generation, but saved practice records store transcript, metrics, and feedback rather than a saved practice audio file.

Known limitations:

• Speech transcription can be wrong, especially with background noise, accents, poor microphones, or interruptions.
• Spoken coaching is preparation support and does not guarantee interview performance or hiring outcomes.

4. Resume, Story, and Profile Extraction

Purpose: Help candidates convert uploaded resumes and guided answers into structured career data and candidate-owned stories.

Primary users:

• Candidates onboarding or improving their career site.

Inputs:

• Candidate-uploaded resumes, profile information, typed answers, and guided story-coach responses.

Outputs:

• Structured profile fields, job history, story drafts, suggested questions, and refinement suggestions.

Human oversight:

• Candidate reviews, edits, publishes, or deletes structured career data.
• Candidate remains responsible for truthfulness and accuracy.

Primary risks:

• Incorrect extraction from a resume.
• Overconfident story rewriting.
• Publishing private or sensitive details by mistake.

Current controls:

• Candidate controls publishing.
• Data retention and privacy pages explain what is collected and retained.
• Account deletion and core profile-data export paths are available.
• AI instructions prohibit adding facts not present in the candidate's answer or source material.

Known limitations:

• Extraction quality depends on input document quality.
• Automatically extracted background is lower confidence than structured candidate-reviewed fields.

5. LinkedIn and Career Copy Assistance

Purpose: Help users draft professional career copy from their own profile and stories.

Primary users:

• Candidates improving their public career narrative.

Inputs:

• Candidate-approved profile information and career stories.

Outputs:

• Draft career copy, LinkedIn-style post text, and positioning language.

Human oversight:

• Candidate reviews and chooses whether to use or publish generated copy.

Primary risks:

• Over-polished copy that implies unsupported achievements.
• Accidentally sharing private or sensitive context.

Current controls:

• Candidate remains the final publisher.
• Accuracy guardrails from candidate data apply to career claims.

Known limitations:

• Generated copy is a draft and should be reviewed before use.