I teach deep learning models to judge things — whether an ad is relevant, whether a message is a scam, whether a post crosses a line — and then make those judgments fast and cheap enough to run at scale. Currently ad relevance; before that, a founding scientist on a content moderation and AI safety team.
The part that actually matters is evaluation. Get the measurement wrong and everything built on top of it is guesswork.
Outside work: part-time musician, runner, recovering electrical engineer.
Posts
Understanding Attention: A Code-First Journey Through Transformers
Build attention mechanisms from scratch in PyTorch. We'll start with raw tensors and progressively build to multi-head attention, explaining every reshape, transpose, and dimension along the way.
Read Post
The 10% You Should Never Automate
Everyone's asking what AI can do. The better question is what you shouldn't let it do. Frameworks for deciding what to automate and what to protect.
Read Post
When Should You Build an AI Agent? A Practical Decision Framework
Practical framework to determine when AI agents make sense for your use case. Learn when to build agents and when simpler approaches like prompt engineering or RAG work better.
Read Post
Mistral 7B on consumer hardware
Run Mistral 7B locally on Mac with Ollama for fast seed data generation. Learn CLI setup, prompt formatting, and downstream parsing to generate thousands of samples on consumer hardware.
Read Post
Finding the right words
Understand how LLMs choose words during generation. Learn temperature, top-k, and top-p sampling strategies to balance coherence, diversity, and task-appropriateness in generated text.
Read Post
Paper Review - Embers of Autoregression
Critical review of LLM limitations in low-probability situations. Explores why AI practitioners should understand autoregressive training pressures before deploying LLMs for tasks requiring precise reasoning or uncommon patterns.
Read Post
Multi-label text classification
Learn to build a multi-label text classifier using DistilBERT with imbalanced classes. Covers binary cross-entropy loss, multi-hot encoding, and practical implementation strategies for handling multiple labels.
Read Post
Library version mismatches declared not safe
Critical lessons on matching Python package versions between model development and inference. Learn about safetensors format advantages and why version mismatches cause production failures.
Read Post
Mining word collocations
Extract common bigrams and trigrams from text using Gensim and NPMI scoring. Learn to mine jargon, phrases, and collocations from customer reviews, feedback, and text corpora.
Read Post
Science Talk: Generative LLMs
Comprehensive introduction to generative LLMs covering basics, training processes, and real-world applications. Slides from talk delivered to 70+ attendees.
Read Post
Projects
Skill Quality Coach
Amazon Alexa announced Skill Quality Coach (SQC), a personalized guide to help skill developers build high-quality skills on Alexa
View Project
Data Science: Analyzing crime stats in Seattle and San Francisco
Analysis of criminal activity periodicity, geospatial distribution by district in R.
View Project
Subscribe
All the latest posts directly in your inbox.