People / HR Data engAnalyticsMLLLM / RAG servicesB2B SaaS Backfill cost −12%

Attrition early warning

3 people · 5 months Data engineer · ML engineer · Me — architecture and delivery

the engagement client data — not shown

heard

"By the time someone resigns I have already lost them. I want to know six weeks earlier, and I do not want a list of names on a dashboard."

real problem

The signals were real but the data was among the most sensitive in the company. Any design that produced individual risk scores visible to managers would have been correctly refused.

system built

A privacy-first aggregation layer: team-level risk with minimum cohort sizes, no individual scores exposed, and a retrieval layer over anonymised exit-interview text that surfaces themes rather than people. HR owns access; managers see their team's theme, not a ranking.

what broke

The first design did produce individual scores. We removed the capability rather than restrict it — a permission is a decision someone can reverse later.

the open rebuild same architecture · public data

dataset

IBM HR Analytics Attrition + public exit-interview text corpus

what the rebuild covers

The cohort-aggregation design and the theme extraction. Deliberately no scoring service: the point of the rebuild is the constraint.

artifacts

NotebookCodeDatasetWrite-upDemo

Open notebook