All of my long-form thoughts on programming, leadership, product design, and more, collected in chronological order.
A brief but complete history of Aerbits, from a frustration with illegal dumping in San Francisco to Oakland's vote to become the first U.S. city to back its aerial-intelligence platform, plus an honest look at how things are going.
How modern transformer models like RT-DETR and YOLOv11, combined with SAM3 segmentation and vision-language models, are revolutionizing waste management from the sky. A complete backend rebuild delivers new capabilities for field crews.
On bridging the gap between experience levels, pattern recognition in engineering leadership, and how teams can learn to see what senior engineers see before it's too late.
Sparrow-1 is an audio-native model that enables AI to respond with natural conversational timing. Instead of using silence-based endpoint detection, it continuously models conversational floor ownership at the frame level, achieving 100% precision with median latency of 55ms—responding at the moment a human listener would.
AI is not an alien visitation. It's a mirror. Drawing on Nietzsche's amor fati, this essay explores what it means to actively affirm the transformation AI represents—not as passive acceptance, but as creative participation in shaping what comes next.
Sparrow-0 is a transformer-based turn-detection model that improves how video agents recognize when speakers finish talking. Using semantic analysis instead of silence-based timing, it achieves response times as fast as 600ms while avoiding interruptions, resulting in 50% higher engagement and 80% better retention.
Developing Sparrow changed my understanding of human communication. By shifting focus from what is said to when someone will speak next, I learned that silence carries meaning, micro-moments reveal intent, and timing is crucial for trust. The dual-lock principle—detecting speech ending then confirming an opening—reduced interruptions by 50% and doubled engagement.
Humans wait about 200ms between conversational turns. Achieving this in video AI requires four pillars: streaming small chunks instead of batches, parallelizing 10 processes, careful resource management, and smart memory handling—moving pointers instead of data saved 1.5 seconds, and GPU memory storage delivered 4x speedup.
How will we leverage great wisdom to create great AI? As a dedicated AI enthusiast, I often find myself reflecting on the profound nature of this technology. I asked ChatGPT4 to help me think through this with help from some of the great thinkers in history, finally landing on quotes from Camus, Acton, and Einstein.
Looking to get the magic of graphql/queries.js,mutations.js,... back?
In ten years, I want to see a world where...
A deep dive into the Aerbits platform architecture.
At each phase of my career, I learned valuable lessons, and here they are chronologically.
In my tenure at Trulia, I played a pivotal role alongside two other engineers in the successful implementation of Trulia's innovative market-based zip-code and impression-based pricing model. This article delves into the technical challenges we encountered and the invaluable lessons we gleaned along the way.
Actively interviewing? It's increasingly likely interviewers are OK with you using ChatGPT to help you answer questions.
In 2016 I returned to Zillow to rewrite the Agent CRM in React and GraphQL. This is the story of how we did it.
How I got into software and how I've used it to build a career and a life.
We were invited to exhibit Aerbits.ai at GenerateSF, a conference hosted for the Asia Pacific Economic Cooperation (APEC) in 2023, in San Francisco...
I've been playing around with the 311 dataset in hopes of understanding what's causing my neighborhood to be so dirty. I've found some interesting trends, and I'm sharing them here.
I spent two days of hacking to accelerate the recovery of San Francisco. Government officials including the Mayor, the City Attorney, and three Supervisors all gave their wish list. Among the top city issues was the issue of street cleanliness.
Nearly a year ago I began the journey of formalizing aerbits's intellectual property. Today, I am proud to announce that the USPTO has issued a patent for our core technology.
At the encouragement of my wife and friends, I spent two days on a "company retreat" working out Aerbit's Q4 Objectives and Key Results.