Notes on AI, delivery, and adoption.
Answer-first essays on getting AI from pilot to production: architecture, enablement, and the agentic systems behind it.
How to spot AI writing before you hit send
A quick, non-technical checklist for catching the giveaway signs of AI-written text in your own drafts: the vocabulary, the lists of three, the flat rhythm, the long dash, and the too-neat ending.
Langdrift: scoring how differently a translation lands
Langdrift compares audio of an original and its translation and returns a drift score: how far apart the two land cognitively, catching accurate but off translations.
Why 'make it sound human' never quite works
AI writing gives itself away on three levels: the characters, the words, and the shape of the whole piece. The deepest one is decided while the text is being written, which is why a clean-up pass at the end cannot fully fix it.
Verso: one JSON source, decks branched per audience
Verso defines presentations as JSON instead of slide documents, so one source can branch per audience and export to PDF, HTML, or PNG cleanly.
AI writing has a tell. I built a free tool that removes it.
Ghostwriter is a free, open-source tool that makes AI-written text read like a person wrote it, by removing the hidden characters, the giveaway words, and the report-shaped structure that flag it as machine-made.
Petrify: removing people from footage, entirely on-device
Petrify detects and removes people from fixed-camera footage while keeping the live timestamp ticking, running fully on your own machine with no cloud upload.
Scope beats autonomy: a contrarian take on AI agents
Open-ended autonomous agents make great demos and poor production systems. Tightly scoped, observable, boring agents are what actually survive real workloads.
Multi-agent workflows a single engineer can reason about
A focused multi-agent system one engineer can reason about beats a sprawling one nobody trusts. Scope each agent, define tools explicitly, and observe everything.
An honest look at AI productivity claims, and how to measure 2-3x
A 2-3x productivity gain is real when you measure build success, cycle time, and effort reduction, not vanity metrics. Here is how to tell the difference.
Evals for delivery teams: trusting AI output without re-checking every line
Trustworthy AI output comes from evals built into the real pipeline, not a benchmark: evaluate what matters, gate at the right point, and keep review habits that scale.
Build a tiny no-code AI workflow that saves you an hour a week
Chain a trigger, an AI step, and an action in a no-code tool like Zapier, Make, or n8n to automate one small repetitive task end to end, no programming needed.
Make yourself unnecessary: what good advisory leaves behind
Good advisory leaves a team stronger, not dependent. Capability and judgement should stay with the team, through co-delivery, explicit reasoning, and grown champions.
Make a full song with AI in ten minutes using Suno
Describe the song you want, pick a style, generate a couple of takes, refine the parts you do not like, then export. Suno makes a full track in minutes.
What predicts whether an AI cohort actually sticks
Across 700+ practitioners, durable AI adoption tracks three predictors: training on their own codebase, role-specific tracks, and internal champions seeded in the team.
Nano Banana, explained: what Google's image model is good at
Nano Banana is the nickname for Google's Gemini image model, great at quick edits and keeping a subject consistent across pictures, with a few known limits.
Singapore Airlines: a website feature in five weeks, down from nine
Embedding GitHub Copilot across Singapore Airlines' delivery lifecycle cut a feature from nine weeks to five, with 95% of the work AI-generated and effort down 60%.
Local vs cloud AI: when to run models on your own machine
Run AI locally for privacy, no per-use cost, and offline use if you have decent hardware, or use cloud AI when you want the most capable models with zero setup.
AI training needs a track per role, not one shared workshop
Engineers, leads, POs, QA, designers, and managers use AI for different work, so one shared workshop gives everyone half-relevant material. Role-specific tracks outperform it.
Replicate: run almost any AI model without the setup
Replicate lets you run open AI models for images, audio, video, and text through a simple interface, with no GPU or environment to set up on your side.
The demo-to-production gap is an organisational problem
A better model does not close the demo-to-production gap. The skills that win a pilot do not generalise, so architecture, guardrails, and behaviour change close it.
Cloning a voice with ElevenLabs, responsibly
You can clone a voice with ElevenLabs in minutes, but only do it with a voice you own or have clear permission to use, never someone else's.
Adoption is a behaviour change, not a tooling rollout
Tools do not change behaviour, practice does. A launch plus a generic workshop spikes then fades, while role-specific practice and champions make adoption durable.
How to spot AI slop and keep your own output good
AI slop is low-effort, generic text with vague filler, endless hedging, and no sources, and you avoid it by adding specifics, your own voice, and real editing.
Prompting is editing, not incantation
The real skill with AI is not finding magic words, it is giving feedback. Draft, say what is wrong, refine, repeat until the answer is right.
Stop role-playing your prompts: what works on modern AI
You no longer need to tell modern AI it is a world-class expert or beg it to think step by step. Give it a clear goal, real context, and the format you want.
The anatomy of a good prompt, and three myths to drop
A good prompt is just five plain parts: a goal, real context, the output format, a few constraints, and an example or two. No magic words required.
AI image generation, made simple: Replicate and Nano Banana
Want AI images without a powerful GPU? Use Nano Banana for quick photo edits, Replicate to try many models in your browser, and a LoRA for a repeatable style.
Taps: turning feed-less websites into real feeds
Neurowire taps are per-host CSS-selector recipes that turn a plain HTML listing page into a real feed, so a site with no RSS becomes a first-class source.
What you should never paste into a chatbot
Assume anything you paste into a chatbot may be stored or used for training, so never share passwords, secrets, other people's data, or confidential work.
Dogfooding @neurowire/core to power this blog's feeds
This site publishes its own Atom, JSON Feed, and NWF feeds from one canonical Neurowire model, because dogfooding your own library is the most honest test it gets.
Understanding LoRAs without the jargon
A LoRA is a small add-on file that teaches an image model one style, character or subject without retraining the whole thing, and you can stack a few.
Nocturne: an open-source Cyberpunk 2077 design system
Nocturne is an open-source design system inspired by Cyberpunk 2077 and Edgerunners: near-black surfaces, one hot neon accent, clipped corners, and machine-voice type, shipped as design tokens, framework-agnostic CSS, and React components.
Building Neurowire: one canonical model, six feed formats
Neurowire treats feeds as a data-modelling problem: one canonical model that every parser produces and every serializer reads, so adding a format is a single serializer.
The everyday AI toolkit: Claude, Suno, ElevenLabs and friends
Match the job to the tool: Claude for writing and thinking, Suno for music, ElevenLabs for voice, and Replicate for running image models like Flux.
The open web forgot about feeds, and that is worth fixing
Feeds let you publish once and let anything read it. As they eroded, following became an account on someone else's platform. Owning what you read is independence.
AI for your actual day: ten small wins most people miss
AI is most useful for small daily tasks like drafting replies, summarising documents, planning trips, and tidying spreadsheets, not just for big flashy projects.
Designing NWF: a feed format 62% smaller than JSON Feed
NWF is Neurowire's native feed format, roughly 62% smaller than the equivalent JSON Feed yet fully round-trippable, using interning, relative links, and date deltas.
How to fact-check an AI answer in 30 seconds
Open the cited source and check it actually says what the AI claims, then cross-check one independent source, paying closest attention to numbers, dates, and names.
Why I optimise for the unremarkable workflow, not the wow demo
A hundred engineers using AI by default beats one jaw-dropping demo. Production compounds on the thousandth unremarkable run, so measure throughput, not wow.
Neurowire Taps Pack: 271 curated sources, 24 themes
The Taps Pack is a ready-made library of 271 vetted sources for Neurowire, grouped into 24 themes from Frontier AI Labs to Food. It's the curation layer ahead of the upcoming Neurowire SaaS.
Neurowire: an open feed engine, now with full docs
Neurowire turns any source, even feed-less websites, into one canonical feed you can serialize to Atom, JSON Feed, Markdown, RSS, or its own compact NWF format. The full documentation site is now live.
Why AI pilots stall before production
AI pilots stall because they optimise for a demo, not for adoption. Production requires architecture, guardrails, and a change in how teams work, not a better model.
What 700 trained engineers taught me about adoption
After training 700+ practitioners, the pattern is clear: adoption is a behaviour change, not a tooling rollout. Role-specific practice and internal champions are what make it stick.