Applied Scientist, Amazon Alexa NLU
- Worked on adopting a multi‑lingual encoder architecture for meeting latency constraints and minimising cost of NLU models.
- Performed A/B testing on the new architecture to expose it to small percentage of traffic to analyse the performance before launch.
- Captured metrics like latency, friction to come up with patches to fix the failures iteratively, also calibrated model confidence to reduce false accepts and false rejects