Skills Can Be Taught. Values Cannot. How TalentOS Built a Hiring AI Around That Idea.

The Problem

Most hiring platforms ask the wrong question.

They match resumes to job descriptions. Keywords to requirements. Experience to credentials. The result is a pipeline full of candidates who look right on paper and fail inside six months.

TalentOS was built on a different premise: skills can be taught. Values cannot.

Rafael Perez, CTO, came to Vivify with a clear mandate. Build an AI platform that evaluates candidates against organizational culture, not just credentials. Something that answers “will this person thrive here?” before the first interview is even scheduled.

No existing tool did it. So they built one from scratch.

The Architecture

Step 1: The Cultural Blueprint

Before any candidate touches the platform, the organization maps its own DNA.

Leadership and HR complete a Values-Vision-Mission survey. The AI synthesizes those responses into a Cultural Blueprint: the organization’s core values, translated into behavioral indicators. Not generic values. Not industry templates. This specific company’s values, in its own language.

That blueprint drives everything downstream. Assessment questions are generated against it. Every candidate is scored against it. The benchmark is custom to each organization, not borrowed from a database of someone else’s culture.

Step 2: The Dual-LLM Validation Loop

This is the part most AI hiring tools skip entirely.

Every assessment question goes through two models before a candidate ever sees it. The first model generates questions based on the Cultural Blueprint. Scenario-based, open-ended, behavioral. “Tell me about a time when…” and “What would you do if…”

The second model independently validates every question. It checks for bias and discriminatory language, ambiguity, redundancy, and cultural sensitivity issues before anything reaches a candidate.

When the second model flags a question, conflict resolution logic kicks in. Minor issues get auto-rewritten. Ambiguous cases get flagged for human review. Clear bias or legal risk triggers automatic rejection and regeneration.

This is LLM-as-judge. One model produces output. A second model audits it before that output reaches a human. It is the quality assurance architecture most AI builds are missing.

Step 3: The Scoring Engine

Candidates respond in narrative form. No multiple choice. No keyword matching.

The scoring engine uses semantic embedding models to evaluate each response. The AI understands meaning, not vocabulary. A candidate who says “I always kept the team informed” scores the same on Transparency as one who says “I maintained open communication channels with all stakeholders.” Same meaning. Different words. Semantic scoring catches both.

Why RAG, Not Fine-Tuning

The Vivify blueprint made a deliberate architectural choice: RAG over fine-tuning. Here is why that call was right.

TalentOS needed to stay current. New roles. New values. New organizational structures. With RAG, updates to the retrieval corpus are near-instantaneous. No retraining cycle. No downtime. The model learns as the organization evolves.

Fine-tuning is the right call when a model needs to learn a new output style or format that prompting alone cannot achieve. For knowledge-heavy, dynamic systems, RAG wins every time.

The Candidate Experience

The platform was designed to be frictionless. Candidates receive a unique URL. No account creation. No login. Mobile-responsive on any device. A progress bar visible throughout, an estimated completion time shown upfront, and session save so a candidate who pauses can return where they left off.

Every moment of friction is a drop-off. This is the same principle Vivify applies to every intake AI built for service businesses. The person who is finally ready to act at 10pm cannot hit a login screen or a dead end. The AI has to capture them in the moment they are ready. Or it loses them permanently.

What Vivify Delivered

Paid Discovery mapped the full architecture before a single line of code was written. That blueprint defined the Cultural Blueprint generation logic, the dual-LLM validation loop, the semantic scoring engine, the HR manager dashboard, and the candidate-facing experience.

The development partners built to spec. Rafael’s team shipped a production platform that evaluates cultural alignment at scale, with a quality assurance layer most AI hiring tools have never considered building.

This is what bespoke AI looks like. Not a product demo layered over a generic tool. A system built around how this organization actually thinks about talent.