Writing
Things I learned by shipping and operating real systems. Not tutorials. Calmly opinionated.
Since Android 10, normal apps can't capture call audio, so Play Store recorders fall back to the microphone and the other person comes out faint. How a Shizuku shell process gets the real call audio without root, the Samsung bugs I had to fix, and a free download.
AI receptionists cost $25 to $500 a month for self-serve apps, about $0.07 to $0.31 a minute if you build your own, and a fraction of a human answering service. Prices from 17 providers, checked on 6 November 2026, with a worked example for a business taking 200 calls a month.
Small businesses in the US and Canada run their customers through a CRM, a form tool, a calendar, an email platform, Zapier, an e-signature app and a payment link, then lose the lead in the gaps between them. The problem we set out to fix, how SpaceBrain works, and how to try it free.
A master key committed to a public repo in May was used 136 days later to steal an org owner's GitHub token and ship a credential-stealing LiteLLM 1.107.0 to PyPI and Docker. What happened, what to do if you installed it, and the practical lessons for AI gateways, agents and API keys.
xAI's Grok Voice 3.0 now tops Artificial Analysis' speech-to-speech index at $0.08 a minute. How it compares with OpenAI's GPT-Live-1 and Google's Gemini 3.8 Live on quality, agent tasks, latency, pricing models, languages and telephony — from someone who runs voice agents in production.
Sierra and Meta's Personal Agent Protocol gives AI agents a front door to your business: a /.well-known/poppy.json file, OAuth sign-in with scopes the user chooses, and a choice of website, APIs or your own agent. What it does, what it doesn't yet, and what to do about it now.
Claude Haiku 5.5 matches GPT-6 Luna to the cent at $0.10 / $0.50 per million tokens. The difference is a pricing step: past 100K tokens, Haiku 5.5 bills the whole request at 5×. Worked numbers for classification, RAG and agent loops.
Google's Gemini now phones businesses for its users — navigating menus, waiting on hold, booking appointments. I build AI receptionists that answer those calls. Here's what changes when both ends of the line are machines.
GPT-6 Astra and Claude Fable 5.1 both cost $10/$50 per million tokens. So do two more tiers below them. The real differences are cache reads, long-context surcharges and the bottom tier. Updated for Claude Haiku 5.5 and Sonnet 5.5.
TypeSafe AI shipped a model that only makes typed decisions — no text generation at all. For the routing hop in a voice agent, that's exactly the right shape. The API, the arithmetic, and the things it genuinely cannot do.
Business-initiated WhatsApp calls are billed per minute in six-second pulses, blocked outright in five countries, and revoked automatically after four rejections. A correction, and what it means for outbound voice agents.
The per-minute anatomy of a production voice agent, why telephony architecture decides your bill, and the WhatsApp trap nobody prices in. With a calculator.
OpenAI's new flagship and Anthropic's both cost $10/$50 per million tokens. For anyone running agents in production, the interesting differences are everywhere else.
Laravel sites on four PHP versions, Next.js apps, Dockerised stacks, databases, and a chat server — on a single machine. Here's how, and when you shouldn't.
Running first-party WebSockets for a real SaaS — nginx quirks, silent disconnects, and why your queue worker is secretly part of your realtime stack.
On a phone call, 800ms of silence is an eternity. Lessons from running voice AI in production.
A production postmortem pattern: why queue workers fail silently, and the Supervisor config that prevents it.
Hard lessons learned running Laravel on constrained infrastructure.
Understanding Expo’s limits before committing long-term.
Patterns and practices for building maintainable API backends.
Tools, configurations, and workflows that make me productive.