Writing
I write when I need to understand something properly. Mostly ML systems, math, and ideas that make more sense once you can poke at them.
Interactive essays
- Why LoRA works: fine-tuning a giant model with a tiny matrix · 9 min readI fine-tuned GPT-2 on movie reviews and found the weight update could be rebuilt from a handful of small pieces. This essay lets you rebuild it yourself, one rank at a time.
- The Determinant: How Much of the Plane Survived? · 10 min readI spent more time than I'd like to admit computing ad − bc without knowing what the number meant. It's the signed area of a parallelogram. Once I saw that, everything else in this topic stopped feeling like a list of rules to memorize.
- What Does a Matrix Actually Do? · 8 min readA 2×2 matrix is four numbers, and those four numbers decide where every point in the plane goes. Drag the basis vectors and watch the grid rotate, stretch, shear, and flatten.
Tutorials Dojo, 2023 to 2024
AWS and machine learning guides I wrote as a technical writer. They open on tutorialsdojo.com.
- Securing Machine Learning Pipelines: Best Practices in Amazon SageMaker · 12 min read · Tutorials DojoWhat SageMaker gives you for security and where it fits in an ML pipeline: encrypting the data, controlling who can access what, and setting up monitoring and logging. (opens in a new tab)
- Training an Image Classification Model with TensorFlow in Amazon SageMaker · 15 min read · Tutorials DojoA small convolutional net trained on MNIST with TensorFlow in SageMaker. Most of the post walks through the training script and the notebook that launches it. (opens in a new tab)
- Train and Deploy a Scikit-Learn Model in Amazon SageMaker · 10 min read · Tutorials DojoA scikit-learn classifier on the Iris dataset, trained in SageMaker, tuned with its hyperparameter search, and deployed to an endpoint. (opens in a new tab)
- Deploying a Serverless Inference Endpoint with Amazon SageMaker · 8 min read · Tutorials DojoAn XGBoost model trained on the abalone dataset and served from a SageMaker serverless endpoint, so you pay per request instead of keeping an instance running. (opens in a new tab)
- How I Prepared for the AWS Cloud Practitioner CLF-C02 Exam as a Data Scientist · 7 min read · Tutorials DojoHow I studied for CLF-C02 coming from data science. I started with a diagnostic practice exam, spent most of my time on the domains I scored worst in, and did hands-on labs alongside the reading. (opens in a new tab)
- Amazon AI Fairness and Explainability with Amazon SageMaker Clarify · 14 min read · Tutorials DojoUsing a crafted loan-approval dataset with bias built in, I trained an XGBoost model and used SageMaker Clarify to see whether it would catch the bias and explain the predictions. (opens in a new tab)
- Automating Binary Classification Model Building with Amazon SageMaker Autopilot · 11 min read · Tutorials DojoHanding a binary classification dataset to SageMaker Autopilot and letting it pick the model. Covers setting up the job, the training modes it offers, and what to look at in the results. (opens in a new tab)
- A Compact Guide to Building Your First DAG with Amazon Managed Workflows for Apache Airflow · 9 min read · Tutorials DojoA small ETL DAG that pulls a CSV, transforms it, and loads the result, running on Amazon MWAA. Covers the S3 layout for DAGs and requirements and creating the Airflow environment. (opens in a new tab)
- Serverless Model Deployment in AWS: Streamlining with Lambda, Docker, and S3 · 13 min read · Tutorials DojoPart two of a deployment series. The model lives in S3, the inference code runs in a Lambda function packaged as a Docker image, and the predictions get written back out. (opens in a new tab)
If something in here is wrong or unclear, send me a note.