Writing
Field notes from production AI.
What actually happens when LLM systems meet real organisations — with the data, the diagrams and the mistakes left in.
In depth for engineers and hiring managers · ELI5 plain-English versions for everyone else — switch at the top of each essay.
The Forward Deployed Engineer: Why Everyone Is Hiring One, When You Actually Need One, and How to Spot a Fake
FDE job postings grew more than tenfold in a year, and the biggest AI companies are committing billions to put engineers inside their customers. The role is a real answer to a real problem — but most companies hiring one have a different problem. A field guide from someone who does the job.
Read the essay →How Close Are We to B-Bots? Ron's Gone Wrong, Five Years On
In 2021 an animated film imagined a tech giant giving every child a robot best friend, out of the box. Five years later the robot's mind is already in our phones, and its body is arriving as $399 ducks that walk, fall over and get back up. The part still missing is the part the film was really about.
The FDE Cookbook: Nine Recipes for Mixing Humans, AI and Deterministic Code
Most of a forward deployed engineer's job is not prompting a model. It is deciding, step by step, which parts of a client's workflow get plain code, which get a model, and which get a person — and wiring the three together so the result survives a Monday morning. A step-by-step recipe book for engineers who want to do this work.
Carole & Tuesday Saw It Coming: What AI Music Can Copy in 2026, and What It Can't
A 2019 anime imagined a Mars where machines write the hits and two teenagers write their own songs anyway. Seven years later, AI tracks are more than half of Deezer's uploads on peak days. The show was right about what that leaves scarce, and the fights it kept in the background are now the main event.
Deploying on a Budget: Vercel + Supabase, Step by Step (and When to Pick Something Else)
A complete walkthrough for shipping a small database-backed web app on Vercel's Hobby plan and Supabase's Free plan for $0 a month, with the two security settings that make it safe, the limits that bite first, and the jobs where another free tier is the better choice. Prices and limits checked on 14 October 2026.
Independent Contractor vs Corporate Employee: The Real Trade-offs Inside Tech Consulting
Going independent looks like a raise. In the Philippines it is closer to a trade: you swap a bundle of statutory benefits, structure and income continuity for rate, autonomy and international reach. A field guide, with a worked break-even calculation, a readiness scorecard and a transition plan, from someone who has worked on both sides.
AI in Public Service: Augment the Caseworker, Never Automate the Verdict
Government AI use cases doubled in a year. The successes share a shape — high-volume, reversible work with a named human in charge — and so do the disasters. A field guide to telling them apart, with the evidence graded honestly and a plan for the Philippines.
Open-Weight LLMs on Windows: When a Quantized Local Model Is a Credible Production Choice
For privacy-bound and cost-sensitive work, a well-quantized open-weight model on an ordinary Windows machine is now a serious option. The model you download matters less than the bits you store it in and the memory budget you fit it into. Here is the evidence, the arithmetic and the setup.
Breaking Into AI: Why Proof Now Beats Paper, and How Students, Fresh Grads and Career Shifters Can Build It
AI skills carry a growing pay premium, degree requirements are falling, and the entry-level door is narrowing at the same time. A field guide to the five realistic paths into AI work, what each one demands, and the 90-day plan that gets you from coursework to evidence.
Productionizing with Git: How GitHub Actions Turns Good Habits into Gates
“Production” is not a server you deploy to. It is the path a change has to travel to get there. This is a practical guide to building that path for an AI or data project with Git and GitHub Actions: CI with an eval gate, approvals and OIDC for deploys, scheduled canaries, and the cheap supply-chain controls that the 2025–2026 attacks showed were not optional.
Backward Deployed Engineering? Maybe It's Not BDE. Maybe It's Called Upstream Engineering.
Forward-deployed engineers carry the product into the customer's world. Somebody has to carry the customer's world back into the product, or every deployment stays a custom project. A deep dive into the economics, mechanics and organisational design of the return path.
ELI5, Intern or Caveman: How to Talk to Coding Agents Without Shipping Code You Don't Understand
Coding agents now speak fluent jargon: tool calls, subagents, compaction, MCP. The way you talk to them, and the way you make them talk back, decides whether you understand what ships. This piece sets out three registers, when to use each, and what the evidence says about understanding, trust and token costs.
AI Governance That Ships: Moving Accountability From the Policy Binder Into the Request Path
Half of organisations using AI have already seen it go wrong, and most governance programmes still live in documents nobody's code reads. Here is what regulators actually require of engineers, the seven controls that turn policy into evidence, and a 90-day plan to build them.
When Is Vibe Coding Okay? The Case Against Perfect Systems for Proofs of Concept
Most proofs of concept never reach production, so polishing their code is usually waste. Vibe coding, which means building by prompt and accepting the AI’s code without reading it, is the right tool when little can go wrong and the code will be thrown away. The professional skill is knowing exactly when to stop.
Stats, Classic ML and Human Review Around LLMs: The Upgrade That Beats a Bigger Model
When an LLM system is too expensive, too error-prone or too hard to trust, the usual reflex is a bigger model. The evidence points somewhere less glamorous: a router, a small classifier, a calibrated threshold, a conformal guarantee and a well-sized human review queue. This article covers how each works, what it buys you and where it breaks.
Model Drift: The Ticking Time Bomb Inside Every “Production” AI System
Nine in ten models in one study got measurably worse after deployment. With hosted LLMs, the model, the serving stack and the harness around it can also change underneath you while your own code stays the same. This is a field guide to the five kinds of drift: how to detect them, set thresholds you can defend, and respond before a customer notices.
Winning Hackathons, and Coaching the Teams That Do: Build Less, Prove More
Judges at the biggest hackathons get two to four minutes per team and score against a handful of near-identical criteria. The teams that win make those minutes easy to score: a sharp user, real data, a demo that works on the path that matters, and a credible next step. A field guide from a four-time podium finisher who now coaches hackathon teams.
Designing with Claude Without the AI Slop: Constraints, Skills and Verification Loops
Ask any model for a landing page and you get the statistical centre of every landing page it has learned from. Better adjectives will not fix that. Encoding taste as constraints, packaging it as Claude Code Skills, and verifying every render with real measurements will.