AI

16 entries

2026

View case study
The City

I saw AI Office showing Claude Code agents as robots at desks, and wanted something that fitted the way I work. In The City, each project folder is a building and each team has a floor of robots, so I can see which ones need me.
- AI
- Architecture
- UX design
Source code
View case study
Typed Diff Questions

A narrow question about a diff can be scored, where a whole code review can't. This was a test of whether a bank of typed questions gives a usable signal.
- AI
- Developer tools
- Research
View case study
Otto

Otto is a voice and text assistant for the Mac, built as a Swift app on top of the voice-os project.
- AI
- macOS
- Swift
- Voice
Source code
View case study
Lumen

Lumen connects an Obsidian vault to a local research backend, so related notes surface while you write and reports come out of what's already there.
- AI
- Obsidian
- Python
Source code
View case study
Synod

Synod reads your journal through a set of value agents and writes their disagreements back into the vault as a bulletin.
- AI
- Obsidian
- TypeScript
Source code
- 26 March 2026 Writing
  Splitting GPUShare into two services
  I split GPUShare’s accounts from the hardware services, then found the installer still pointed at a service I’d removed.
  - AI
  - Architecture
  - Economics
  - Self-hosted
Two layers: an always-on cloud layer holding accounts, ledger and invoices, and a home GPU host running Ollama and Blender that connects outbound through a tunnel with no open ports.
Idle GPU, billed at cost

Two services, one trust boundary

Always on

Accounts · ledger · invoices Reachable with your PC off

Outbound tunnel

No open ports The host connects out; nothing connects in

Your machine

Ollama · Blender Only while powered on

kWh rate × GPU watts = price. No markup.

Read the story
GPUShare

My desktop GPU sits idle most of the day. I built a way for friends to share it for local AI and rendering, with the limits of a home machine visible.
- AI
- Architecture
- Economics
- Front-end development
- Self-hosted
- UX design
Source code
- 21 March 2026 Writing
  Sharing my GPU with friends
  I built GPUShare around my desktop GPU. The accounts, routing and cost estimates ended up needing as much explanation as the hardware.
  - AI
  - Architecture
  - Economics
  - Self-hosted
- 8 February 2026 Writing Report
  Anomaly (Out-of-Distribution) Detection for Electrical Distribution Networks using Discrete Diffusion Methods
  The paper behind Grid Diffusion AD: discrete diffusion over quantised magnetic-field windows, trained on normal data only, scored against a false-alarm budget. The best configuration separated faults from normal weakly, and the paper reports that instead of rounding it up.
  - AI
  - Machine learning
  - Research
  - Time series
- 20 January 2026 Writing
  Putting my Apple Health data in one place
  I brought my nutrition and activity data out of Apple Health and into Postgres, so I could compare the daily and weekly trends.
  - AI
  - Architecture
  - Data pipelines
View case study
HealthAgent

My nutrition and activity lived in different apps. I brought their Apple Health exports together so I could see the trends in one place.
- AI
- Architecture
- Data pipelines
- Front-end development
Source code

2025

View case study
Grid Diffusion AD

Diffusion models trained on normal magnetic-field signals, scored on whether they notice a fault they've never seen.
- AI
- Machine learning
- Research
- Time series
View case study
Pinterest Aesthetic Analysis

I used CLIP embeddings to group the images on a Pinterest board, then kept the snapshots so one month's themes could be compared with the next.
- AI
- Computer vision
- Python
Source code
- 9 June 2025 Writing Report
  Wildfire Analysis with PySpark
  I used PySpark to train and compare models on 1.88 million US wildfire records. The report covers the features, training runs and what happened when I weighted the rare classes.
  - AI
  - Data pipelines
Recorded random-forest results: class G recall rose from zero to 0.7518 with weighting; class G precision was 0.0099 and overall accuracy was 32 percent.
PySpark experiment

1.88M records

Class G recall

0 → 0.7518 With class weighting

Weighted class G precision

0.0099

Weighted overall accuracy

32%

Recorded random-forest results. Class weighting finds more rare fires, with many more false positives.

View case study
Wildfire PySpark

I worked with 1.88 million US wildfire records for a university project, and got into how a model could look accurate while missing the larger fires.
- AI
- Data pipelines

2023

View case study
ArticulateAI

One of my early voice experiments: record a clip, transcribe it with Whisper, send the transcript to GPT with a prompt, and optionally speak the reply.
- AI
- Python
- Voice
Source code