CASE 03 · 2025 · Co-builder · Multi-agent AI
SENTINEL
Privacy-first enterprise wellbeing analytics with a 3-agent AI orchestra.
SYSTEM FLOW
THE PROBLEM
Burnout is a business risk HR can't see until it's expensive. Any solution had to be privacy-first — anonymized signals, not surveillance — and answer plain-language questions, not just dashboards.
3 specialized AI agents in one system
HOW IT'S BUILT
Co-built a FastAPI + React platform scoring burnout risk from anonymized interaction patterns, with a graph-based module for team-collaboration insight. Implemented a 3-agent orchestration system — burnout scoring, talent discovery, team health — enabling natural-language queries over workforce analytics.
IMPACT
- 3 specialized AI agents in one system
- Graph analysis for team-collaboration insight
- Privacy-first: anonymized patterns only
- Natural-language queries over analytics
WHAT IF …
Ask the project a different question. The architecture has to defend itself.
What if you used a single LLM instead of 3 agents?
One model doing scoring + talent + health would blur the evidence trails and make audits impossible — and hallucinate team conclusions. The three-agent split keeps each system small, testable, and accountable; the orchestrator only routes. More moving parts, but each one is honest.
What if the board demanded per-person scores?
Refuse, architecturally. The whole design is privacy-first: anonymized patterns only, no identity in the pipeline. I'd show them cohort-level risk trends, team-collaboration graphs, and department aggregates — the same insight without the surveillance.
DIVE DEEPER
Built it — now the descent. Each question opens the next layer: why, why this architecture, what broke, what I'd change.
01Why three agents?
Burnout risk, talent discovery and team health are three different questions with different data shapes. One model would blur them; three specialists stay honest.
02Why privacy-first from day one?
Wellbeing analytics dies the moment it feels like surveillance. Anonymized patterns only — no names attached to scores.
03What went wrong?
Natural-language queries were slow until I pre-aggregated graph features per team. Latency dropped from seconds to ~300ms.
04What would I do now?
Add temporal decay to signals and a human-in-the-loop review queue before any score leaves the system.
THE REPO, INSIDE
Not a screenshot — a live iframe pulling this repo's README straight from the CDN.
THIS PROJECT LIVES IN A PRIVATE WORKSPACE