2026-03-09
An LLM persona simulation tool for predicting job-person fit. Personas constructed from Big Five personality traits and work values are dropped into realistic workplace scenarios. Their responses are automatically scored across a 9-dimension behavioral rubric, producing a predictive fit profile.
Traditional job-person fit assessment relies on self-report questionnaires and interviews — slow, expensive, and prone to social desirability bias. The question driving this research: can LLM persona simulation predict behavioral patterns under occupational stress before real-world deployment?
1. Agent Generation. You set Big Five personality traits and work values on sliders. Gemini generates a narrative context describing the agent's likely workplace behavior. An NPC colleague network — 11 fixed characters across 5 departments — is initialized with dynamic rapport scores and relationship types.
2. Scenario Simulation. 25 workplace scenarios across 5 job roles (Sales, Dev, Planning, HR, Research), each with 5 stress situations. The agent runs through a multi-step agentic loop using Gemini Function Calling — autonomously selecting tools, integrating results, and producing a structured response: immediate emotion, concrete action, inner monologue.
3. Tool Use. The agent has four tools at its disposal:
| Tool | What It Does |
|---|---|
| search_company_policy | RAG retrieval over internal policy documents |
| ask_colleague | Traverses the NPC graph, finds the best colleague by rapport and expertise |
| check_past_cases | RAG retrieval over a library of past success/failure cases |
| request_data | Pulls internal organizational data — sales figures, schedules, satisfaction metrics |
4. Scoring. Each response is scored across 9 behavioral dimensions on a 1–5 BARS-anchored scale:
Behavioral: Adaptive Capacity, Achievement Motivation, Job Involvement, Self-Efficacy, Attitude
Tool-use: Tool Selection Quality, Information Seeking Behavior, Collaboration Pattern, Reasoning Depth
Scoring supports LLM auto-scoring and independent dual-rater manual entry, with Cohen's Weighted Kappa computed in real time.
The same agent × scenario combinations run under 4 prompt strategies to measure their effects:
| Strategy | Description |
|---|---|
| Baseline | Direct persona + scenario prompt with tools enabled |
| Chain-of-Thought | Explicit 5-step reasoning before action |
| Persona Anchoring | Repeatedly reinforces behavioral anchors for consistency |
| Tool-Encouraged | Explicitly encourages tool use and step-by-step information gathering |
Statistical analysis includes Cohen's Weighted Kappa, Krippendorff's Alpha, and Cohen's d for effect sizes between strategies.
React 19 + TypeScript + Vite on the frontend, Express / Vercel Serverless on the backend, Google Gemini (gemini-2.0-flash) with Function Calling for the AI layer. State management with Zustand, visualizations with Recharts and HTML5 Canvas.
HCI collaboration with Cornell HCI Lab. MOU signed with POSCO HR for deployment in new-hire assessment workflows. Built on the SPeCtrum framework.