ML Software Engineer
at Amazon
I work on interpretability and personalization at scale, turning opaque embeddings into features people can read, and running them in production for hundreds of millions of customers.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
class SamKapusta(AmazonEmployee):
"""Interpretability + personalization, in production."""
def machine_learning(self):
# Sparse autoencoders -> audience segments a human can read.
# Winning model now backs the Amazon Ads advertiser product.
return SAE.train(features=4_000, customers=330_000_000)
def software_development(self):
# Led a team of 4. Daily inference, monitored + alarmed.
return StepFunctions() >> SageMaker() >> Redshift()
def data_engineering(self):
# 50M+ daily events -> per-customer sequences.
return stream.vend(to=["risk", "ads", "logistics"])
>>> sae = SAE.load("llm.residual_stream") # a trained sparse autoencoder
>>> sae.decompose("the Golden Gate Bridge at sunset").top(4)
Golden Gate Bridge0.94
San Francisco0.71
suspension bridges0.52
dusk / warm light0.38