Synchrony Hackathon2026Hackathon
Cadence
Behavioural fraud scoring: does the applicant know their own identity?
Real-time fraud scoring for digital lending. Your own name and date of birth are practised motor sequences; a stranger's are not. Cadence compares keystroke rhythm on identity fields with fields everyone reads off a document, inside one session, so each applicant is their own control. It never reads field contents.
livelast push
~/cadence
Project Telemetry
Humans showing effect
148 / 150
Speed-group spread
6.8% vs 87.6%
Channels
4
Auto-declines
0
How it works
~/cadence/architecture · step 1/3
01
keystroke SDK (no field contents)
02
identity vs document rhythm
03
4 independent channels
04
corroborate + friction budget
05
human review
01Validated on two public keystroke corpora (CMU, KeyRecs): 148 of 150 real people type practised strings faster and more evenly.
02A population baseline mostly detects slowness (87.6% spread across speed groups); the within-session ratio holds at 6.8%.
03Four independent channels (timing, physical, provenance, structural) cross-check each other; behavioural evidence never auto-declines, it routes to a human.
Project Info
DomainFraud Detection
RoleHackathon build
Year2026
StatusHackathon
Tech Stack
PythonFastAPIscikit-learnpgvectorReactVanilla JS SDK