Conversational AI Platform for Employee Welfare March 2025
An AI-driven system to monitor employee well-being through behavioral signals and automated insights.
Most "employee well-being" tools amount to an annual survey nobody fills out honestly. This project took a different angle: an AI-driven system that infers well-being continuously from behavioral signals instead of asking people to self-report on a schedule, and surfaces insights automatically instead of burying them in a dashboard nobody opens.
How it actually works
Data from 1,000+ users flows through a multi-stage pipeline built on FastAPI, orchestrated by scheduled CRON jobs that continuously recompute behavioral metrics rather than running as a one-off batch job. On top of that:
- Risk scoring: high-risk users are flagged using EMA-weighted scores (exponential moving averages, so a single bad day doesn't dominate the signal) combined with ensemble models, which together improved flagging accuracy by 42% over a naive threshold-based approach.
- Conversational layer: an LLM-powered chatbot, orchestrated through LangChain pipelines, generates context-aware summaries and check-ins rather than generic ones - the system actually reasons about why a user's signals shifted before it says anything to them.
The interesting engineering problem here wasn't the LLM calls themselves - it was making the pipeline trustworthy enough that "AI flagged this person as at-risk" is a claim worth acting on, not just a number that looks impressive in a demo.