Writing

Mostly things I built and what broke. 28 posts since 2017.

2026

30 Jul 2026

The ESP32 pinout cheat sheet I actually wanted

Every pinout diagram tells you what each pin can do. Almost none of them lead with what will stop your board booting, or which pins go dead the moment you turn on Wi-Fi. This is the reference I ended up writing for myself.

esp32
hardware
reference
embedded

27 Jul 2026 Featured

I have been building the same thing for ten years

A JSON layout engine, a schema-to-code generator, two widget frameworks, and now an LLM gateway. Four companies, four stacks, one idea I did not notice until I wrote them down in order.

infrastructure
career
android
llm

26 Jul 2026

The query said the data was dead. Twice, it was wrong.

I was auditing a production database to find what was safe to delete. Two of my findings said "dead, safe to remove" and both were artefacts of how I was measuring.

firestore
databases
production
debugging

24 Jul 2026

Your LLM gateway routes. Mine bills.

Routing between model providers is the commoditised half of the problem. The half nobody wants to build is knowing whose money you just spent — and it is the half that decides whether your AI product has a business model.

llm
infrastructure
billing
gateway

22 Jul 2026

What it costs you when one user writes an infinite loop

Nobody has to be malicious. A retry policy, an agent that does not know when to stop, and a for-loop over a spreadsheet are enough. Here is the arithmetic, and what actually stops it.

llm
infrastructure
billing
reliability

20 Jul 2026 Featured

I built a pocket AI companion out of a ₹400 microcontroller

An ESP32 with an animated face, a microphone, a speaker and a wireless camera eye. You hold a button and talk to it, and it can call tools that move real hardware.

esp32
hardware
ai-agents
buddy

17 Jul 2026

I designed a PCB by talking to an agent

KiCad driven through an MCP server, autorouted, ground poured, DRC clean at zero errors. Also: the four things it got wrong, which are the interesting part.

esp32
hardware
ai-agents
kicad
mcp

15 Jul 2026

The microcontroller is a terminal. Your PC is the brain.

Version one ran everything on the ESP32 and was permanently out of memory. Version two moved the thinking to a proxy and got unbounded speech, streaming replies and swappable models for free.

esp32
hardware
architecture
ai-agents

13 Jul 2026

The flicker was the radar

A desk instrument started flickering on every screen. The cause was a sensor that had nothing to do with drawing, running on a schedule that had everything to do with it.

esp32
hardware
debugging
firmware

13 Jul 2026

Two heartbeats killed my WebSocket every thirty seconds

A healthy link that dropped like clockwork, and a device that would happily retry a dead IP address forever. Two connection bugs where the fix was removing something, not adding it.

esp32
debugging
websockets
networking

20 Jun 2026

I gave an agent a CAD program and asked for LED panels

Modular WS2812B panels, designed by describing constraints instead of drawing them. The interesting part is which half of the design work actually transferred.

hardware
cad
ai-agents
3d-printing
mcp

14 Jun 2026

Telegram was the right UI until I was sitting next to the machine

Phoebe runs on hardware in my house and I talk to it on Telegram. That is excellent from a train and slightly absurd from the chair in front of it.

ai-agents
phoebe
telegram
self-hosted

14 May 2026

Debugging a dead motor driver with a multimeter in one hand and an agent in the other

Four days, three power supplies, two motor drivers and one chip that turned out to be dead. The agent could run code on the robot. It could not see the robot. That division of labour turned out to be the whole trick.

robotics
hardware
ai-agents
raspberry-pi

23 Mar 2026

"OpenAI-compatible" is a contract, not a format

I renamed an environment variable to the accurate name and it silently rerouted my traffic to another company. Four days of small breakages, and what they say about what compatibility actually means.

llm
gateway
api-design
mume

4 Mar 2026

Thirty thousand moving objects and one main thread

A long weekend spent putting every satellite and every aircraft on a 3D globe in a browser. The globe was the easy part. Everything after it was a fight for the main thread.

frontend
performance
webgl
dataviz

27 Feb 2026

The gap between the answer and the bill

You cannot price an LLM call until after it has finished. That one fact decides the shape of every metering system you will ever build on top of one.

llm
billing
infrastructure
mume

25 Feb 2026

Cancelling a stream doesn't cancel the bill

The user closes the tab. Your server writes into a socket nobody is reading. The provider keeps generating, and keeps charging. Nothing errors, which is the problem.

llm
streaming
infrastructure
billing

24 Feb 2026 Featured

I gave an old Raspberry Pi 3 full access to my GitHub. Here's what happened in 48 hours.

A Raspberry Pi 3 Model B with 906MB of RAM, an autonomous TypeScript agent with 855 runtime-loaded skills, and two days of unsupervised access to my GitHub org.

ai-agents
raspberry-pi
typescript
phoebe

23 Feb 2026

Why are we innovating here? We already solved this.

The night Phoebe was built: one abandoned framework, one teardown, and the sentence that turned the evening around.

ai-agents
phoebe
raspberry-pi
telegram

18 Feb 2026

Ten rounds on a rounded corner

An agent can write the CSS. It cannot look at the screen. That one asymmetry decides which parts of interface work go fast and which turn into a loop you have to break yourself.

design
ai-agents
frontend
mume

13 Feb 2026

Nine weeks from first commit to taking money

The commit log of an LLM gateway going from nothing to production, and what the order of it says about which parts are actually hard.

llm
infrastructure
mume
solo

10 Feb 2026

Rate limiting an LLM API is not rate limiting a REST API

Requests are the wrong unit and everyone uses them anyway, including me. Here is why the right unit is so hard to enforce, and what to do until you can.

llm
api
infrastructure
mume

2025

29 Nov 2025

Mume Gateway: one endpoint, 500+ models

Point the OpenAI SDK at a different base URL and you have access to models from OpenAI, Anthropic, Google, Mistral and others — with per-user credits, quotas and billing behind it.