Radical transparency
Why we build, and how
We would rather say the true thing than polish it away. This page covers why TalentProof exists, the principles we steer by, the stack we run, the AI we use, and why assessment results are decision support you can inspect, not a black box hire score.
Why we build this
Hiring teams waste live time on resumes that look strong and interviews that fall apart. In an AI-cheating era that problem got louder. TalentProof is a pre-interview verification layer: one intake link, async filter, then a short proctored step that produces a verified TalentProof scorecard you can stand behind. We are not trying to replace your verbal screen or your ATS. We sit in front of the expensive part.
Principles we steer by
- Defensibility where stakes are high (client submit, live interview), not moral homework about "honesty."
- Align candidate, recruiter, and (for agencies) client around lower delivery risk on the next step.
- Narrow wedge first. Earn the right to bigger platform stories later.
- Same dual-primary buyers in product and outbound: small US staffing desks and lean startup / Series A hiring. Mid-market inbound only.
- Paper Tigers / AI-inflated resumes are the wedge. We frame outcomes (time and trust), not surveillance theater for candidates.
Stack
Product app: Next.js (App Router). Auth and data: Firebase Auth and Firestore. Hosting: Firebase App Hosting on Google Cloud. Sensitive fields are encrypted before storage. Details: data security · privacy.
AI we use
Assessment generation, scoring assistance, match summaries, resume flags, practice interviews, and copilots run through Google Genkit on Vertex AI Gemini in production (default gemini-2.5-flash, with Pro and preview models available for harder jobs). Local development can use a Gemini API key. We do not train public models on your personal data unless you opt in. See privacy.
Why trust the results?
Fair question. A model can sound confident and still be wrong. Here is what we actually claim, and what we do not.
Who judges the responses?
For open responses, an LLM scores against the role: the job description, the question, and a structured rubric for that assessment format. It is not guessing hire or no-hire from a resume vibe. Rapid fire uses answer keys. Hands-on and Conversation have their own rubrics. Details by format: scoring methodology.
What makes a scorecard "verified"?
Proctoring. Rules are disclosed up front. Session signals (focus loss, paste bursts, timing, face presence when camera is on, and similar) attach integrity context to the scorecard. That is a defensible signal before live time. It is not a promise that cheating is impossible. Bypass paths exist. Recruiters still read the Proctoring tab.
Who decides who advances?
A recruiter. AI drafts questions and scores work product. Job match and resume flags are advisory call prep. They do not reject anyone on their own. Pre-screening knockouts and criteria you enable on a role can filter before assessment when you turn that on. Screening behavior follows what you configure on each role. No platform-wide letter grades (A/B/C/D) or hidden AI cutoffs. Full rules: trust & scoring.
Why trust the judge model at all?
Because the input is constrained. Same role brief, same questions, same session context, structured output, and a human who can open the transcript and disagree. We score demonstrated work under proctoring, not identity, pedigree, school, or name. We do not claim bias-free or cheat-proof. We claim inspectable decision support you can defend on a client call or to a hiring manager.