Employee Communication Simulator

Leadership communications often fail not because the message is empty, but because different audiences infer different intent, fairness, and personal risk from the same text. A message that sounds reassuring to a senior leader might feel threatening to a mid-level contractor in a different geography.

I wanted to build a tool that could simulate those differences before a message was sent, then package the output into something executives could act on. Not just raw data. Framed analysis with clear scenario ranges.

What started as a lightweight script for testing employee reactions evolved into a multi-component product: a weighted simulation engine, a multi-provider LLM layer, a Monte Carlo scenario model, a React dashboard, durable async job orchestration, saved-run retrieval, and automated executive reporting in PDF and PPTX formats.

The system models 60 weighted personas representing a 450-person IT organization — spanning employees vs. contractors, US vs. India, leaders vs. ICs, and principal architects. Three simulation modes (heuristic, LLM-backed, Monte Carlo) give different tradeoffs between speed, cost, and depth.

The stack was deliberately lightweight where possible and specialized only where the problem demanded it. The system relies heavily on plain Python with filesystem-based persistence, React for the interactive surface, and a provider abstraction layer across four LLM backends.

Weighted personas over flat personas. Expanded from 15 generic archetypes to 60 weighted personas representing real organizational proportions. Summary metrics now represent a simulated org, not disconnected characters. That distinction is critical for executive trust.

Kept heuristic mode alongside LLMs. Not every run needs hosted inference. Preserving a fast, low-cost simulation path made the tool practical for iterative message drafting, with LLM mode available when depth mattered more than speed.

Saved outputs as first-class product objects. Run listing, retrieval, and reloading were built into the experience, not treated as throwaway files. Executive reporting operates off saved runs. That is what makes a tool credible as a decision-support system rather than a demo.

Durable job lifecycle over request-response. The most consequential shift was moving from "one request, one response" to long-running jobs with persistent state, checkpoints, and explicit failure modes. Once Monte Carlo runs, LLM calls, and background tasks entered the picture, durability became a product requirement.

This project demonstrates the ability to take an ambiguous problem, build the first version quickly, recognize when the original architecture stops being good enough, and evolve the system into something more resilient, interpretable, and useful. Solo, end-to-end, spanning backend, frontend, AI integration, simulation design, and executive reporting.

It shows systems thinking, product design under real constraints, operational hardening of AI features, and the ability to translate analytics into decisions that leaders can act on.