Salome Wairimu

About

I'm a senior software engineer and tech lead on Google Docs AI, where I work on AI quality: building evaluations, iterating on prompts and agent skills against them, and building the infrastructure that agentic features run on. I've been at Google since 2020.

The question I care most about is whether a change actually made a model or agent better, and why. That means evals that measure what users care about, and experiments that separate a real improvement from noise or a lucky configuration.

Alongside work, I'm doing graduate coursework in machine learning at Georgia Tech, with a focus on deep learning, reinforcement learning, and how models hold up under distribution shift.

I studied computer science at Northwestern, where I did HCI research.

Selected work

• Do vision transformers transfer better than CNNs across domains?

  Comparing ViT-Small and ResNet-50 on DomainNet when moving from real photos to sketches, paintings, and doodles. Part of a team transfer-learning project.

  Deep Learning (CS 7643), Georgia Tech · in progress, Fall 2026

• When does SARSA beat Q-learning?

  Value iteration, policy iteration, SARSA, and Q-learning on Blackjack and CartPole. The main finding: SARSA's edge over Q-learning came from the exploration schedule, not the algorithm itself.

  Machine Learning (CS 7641), Georgia Tech

• Unsupervised learning and optimization studies

  Clustering and dimensionality reduction experiments, and a comparison of randomized optimization against Adam-family optimizers for training networks.

  Machine Learning (CS 7641), Georgia Tech

Writing

Notes on ML and agentic AI are coming here. Earlier essays from Northwestern are below.

Elsewhere

• GitHub
• LinkedIn