All work
Anomaly detection
Kestrel Space
2023
Anomaly detection over satellite telemetry that flags failing parts days before an operator would notice.
Scroll to resolve
4.2 days
Earlier failure warnings
Client
Kestrel Space
Role
ML engineer
Stack
JAX · Time-series transformers · Airflow
Year
2023
The problem
Operators watched 3,000 telemetry channels per satellite and mostly learned about failures after they happened.
What I built
I trained a forecasting model per subsystem and alerted on the gap between prediction and reality, with a ranked explanation of which channels drifted first.
The result
Failing components are now flagged 4.2 days earlier on average, with fewer than one false alarm a week.