We built a privacy-first platform that spots burnout risk across engineering teams from de-identified work data and anonymous surveys, then briefs managers so they can act early. Staff turnover fell 24%.
Industry: HR TechService: AI & Automation
Overview
Client
Refuge Agency
Scale-up with 800+ engineers across 4 time zones
The challenge
Strong engineers were resigning with no warning, often straight after good performance reviews. Engagement surveys ran once a quarter, and few technical staff bothered to fill them in.
Each resignation took system knowledge with it and cost a great deal to replace. The client had tried people analytics before, but engineers refused it because they could not see how their personal details were removed.
When senior engineers left, projects slipped by 3 months and the client lost important knowledge of how its systems were designed.
Hiring, onboarding and ramping up a new senior engineer cost $250k+ on average, money that came straight out of the R&D budget.
The engineering committee had rejected earlier people-analytics tools because they did not show clearly how identifying details were removed.
The solution
We built a secure proxy that strips personal details and message content from Jira, GitHub and Slack events. It keeps only timing and engagement data, such as how often people commit code and how quickly they answer pull-request reviews.
A transformer-based AI model compares anonymous pulse-survey sentiment with engineering speed benchmarks and flags groups of engineers at high risk of burnout.
Managers receive automatic briefs with team-level insights and a suggested script for a stay conversation. Every step from alert to conversation is logged, so each intervention can be traced.
Key features
A privacy proxy strips personal details from Jira, GitHub and Slack events and applies differential privacy.
Business impact
Zero personal data exposed
A model compares anonymous sentiment with engineering velocity to flag burnout risk and give managers coaching tips.
Business impact
24% lower churn
Statistical noise is added to the data, so no individual engineer can be identified, even if the database is breached.
Business impact
100% developer acceptance
AI workflows write personalised coaching tips for each manager, based on the burnout signals found in their own team.
Business impact
92% brief adoption
The system spots sudden jumps in pull-request response times and quiet spells in GitHub activity, two early signs that an engineer is pulling away.
Business impact
Zero false flags
Before and after
Before
Quarterly pulse surveys
Manual self-reporting, prone to bias, with only 40% of staff taking part.
After
Everyday work patterns
Signals collected passively and continuously from the development tools engineers already use.
Before
Reading messages
Opt-in manual reviews of Slack or email content, which hurt trust and individual privacy.
After
Hashed patterns only
Models work on hashed data patterns and never see what anyone wrote.
Before
Exit interviews
Insight arrived only after the engineer had already decided to leave.
After
Early intervention
Alerts arrive 30-45 days before at-risk behaviour usually ends in a resignation.
Results
24% Churn Reduction
Early conversations kept engineers who, in the past, would have resigned without warning.
Before: High ChurnAfter: 24% Lower
88% Accuracy
Reliable flags meant managers could spend their time on the teams that needed it most.
Before: 40%After: 88%
$720k First-Quarter ROI
Turnover had cost the business more than $3M in replacement costs. In the first quarter, lower churn kept $720k in the R&D budget.
Before: $3M+ LostAfter: $720k Saved
Testimonial
“With 800+ engineers in 4 time zones, burnout was hard to see early. The model Coretus built now flags at-risk teams at 88% accuracy, up from 40%. Churn is 24% lower, and no personal data has been exposed.”
Key results
Trust and control
Every workflow uses k-anonymity and differential privacy, so individual identities are never exposed.
Aligned with GDPREvery two weeks, the models are checked for gender and racial bias in engagement scores, so retention support is offered fairly.
Bias checkedThe cloud pipeline handles 1M+ development events a week and processes each one in under a second.
Cloud nativeThe client owns 100% of the retention models and privacy proxies once the project is complete.
100% ownedFaster delivery
Starting from these modules saved about 10 weeks of setup, which we spent on the client's own data sources and on tuning the manager briefs.
Removed identifying details from Jira, GitHub and Slack events before any data reached the models.
Gave us a trained starting point for scoring anonymous survey answers, which we then tuned for engineering teams.
Supplied the dashboards and alerts that show team-level burnout risk to People Operations and managers.
Measured the replacement costs avoided, which is how the first-quarter saving of $720k was calculated.
We built a proxy that strips personal details and message content from Jira, GitHub and Slack events before any model sees them. Only timing and engagement signals remain, such as commit frequency and pull-request review times. Differential privacy and k-anonymity add another layer, so no single engineer can be identified, even if the database is breached.
The platform cut engineering churn by 24% and predicts turnover with an 88% F1 score. Managers now get alerts 30-45 days before an at-risk engineer is likely to resign, which leaves time for a stay conversation. Lower churn kept $720k of replacement costs in the R&D budget in the first quarter.
Each manager gets a brief with team-level insights and a suggested script for a stay conversation. The coaching tips are based on the signals found in that team, such as slower pull-request reviews or a drop in GitHub activity. Adoption of the briefs reached 92%, and every intervention is logged so it can be traced.
If good engineers are leaving without warning, we can help you spot the early signs in your own work data without reading anyone's messages.