Revolutionary by design
PoX AI turns live candidate behavior into structured proof that hiring teams can compare, calibrate, and review. The stack combines agent simulation, model reasoning, and pattern-based scoring without hiding the evidence behind a generic number.
Why simulation changes the signal
A good simulation should reveal behavior that a normal interview hides. In sales, customer success, recruiting, support, and operational roles, performance depends on how a person responds to incomplete information, conflicting incentives, and live feedback.
Traditional interviews tend to compress those conditions into hypothetical questions. The candidate explains what they would do, the interviewer interprets the answer, and the organization later discovers whether the answer predicted work.
PoX AI changes the unit of assessment from an answer to an interaction. The candidate is asked to perform inside a simulated business situation: qualify a buyer, handle objections, protect trust, explain tradeoffs, prioritize next steps, or coordinate a decision among multiple stakeholders.
Each exchange becomes a trace. The transcript shows what was said. The agent state shows what pressure was introduced. The voice pattern record shows how delivery changed when the scenario became more difficult.
Architecture
The upper layer produces the simulated world and buyer committee. The middle layer performs model reasoning under controlled prompts and state constraints. The lower layer records speech, extracts patterns, and maps observed behavior to a rubric.
Every scored claim should point back to a transcript segment, scenario event, or voice-pattern observation. When a reviewer disagrees with a score, they can inspect the evidence rather than only seeing a final number.
Inside the stack
The scoring layer converts live behavior into structured observations. Speech-to-text, text-to-speech, and pattern extraction create a synchronized record of what was said, how it was said, and whether the candidate demonstrated the target skill.
Frontier language models provide the adaptive reasoning layer. They operate as controlled personas with scenario constraints, rubric boundaries, and conversation-state memory.
A multi-persona buyer committee creates pressure, ambiguity, interruptions, and conflicting priorities. The candidate is evaluated against how they discover information, manage stakeholder tension, and turn a simulated conversation into an evidence-bearing interaction.
Every call and meeting is embedded with Gemini Embedding 2 and indexed for semantic retrieval. Top-performing conversations become a searchable corpus that surfaces the patterns behind best-seller outcomes.
The scorecard surfaces concrete behavior in reviewer language. A useful AI assessment should make it easier for humans to compare what happened, not harder by hiding the judgment behind a generic score.
What it catchesQuestion sequencing, explicit hypothesis testing, and the ratio of follow-up questions to generic prompts.
What it meansHigh scores indicate that the candidate forms a useful model of the buyer instead of performing a scripted checklist.
What it catchesTreatment of conflicting incentives across finance, technology, procurement, and end-user personas.
What it meansHigh scores indicate that the candidate can hold multiple constraints in the same conversation without flattening them into one answer.
What it catchesSpecificity of claims, requested proof, acknowledgment of uncertainty, and correction after challenge.
What it meansHigh scores indicate that the candidate anchors the interaction in observable facts rather than confident but unsupported statements.
What it catchesPace, pause placement, filler density, interruption recovery, and confidence stability during difficult turns.
What it meansHigh scores indicate that pressure changes the candidate's behavior less than it changes the simulated environment.
Searchable corpus
Every call and meeting in the platform is embedded with Gemini Embedding 2 the moment it ends. Transcripts, scoring evidence, and voice-pattern observations become vectors in a shared index, not isolated recordings.
Top-seller conversations are tagged at the source, so a query like "how the best reps handle a procurement objection" returns the actual moments behind the outcome instead of a summary.
Employers can search prospects across the full PrompX database and hire the best, comparing candidate simulations against the closest reference calls from real winners. Scoring stops being abstract and starts pointing at the patterns that produced revenue.
Fairness controls
The stack must be calibrated with care. Voice features can be useful signals of pressure response, but they should not become proxies for accent, disability, background, or communication style unrelated to job performance. For that reason, delivery features should be used as supporting evidence and should remain inspectable by reviewers.
Agent simulations also require calibration. If a persona is too hostile, the task measures stress tolerance more than job skill. If a persona is too agreeable, the task becomes a scripted demo.
The strongest version of PoX AI is not autonomous judgment. It is a repeatable evaluation layer that gives human reviewers better evidence than a conventional interview.