Work
ProjectsCase studies, shipped work, and prototypes.Open SourceGitHub, npm, PyPI, and public packages.Hackathons & CompetitionsHackathons, contests, and rapid-build work.Depth
JourneyCareer timeline, roles, and education.AchievementsScholarships, rankings, and milestones.Courses & CertificationsStructured learning, courses, and credentials.SkillsTechnical capability map.This website · 02
Browse this website posts across architecture, delivery, product systems, and the public writing footprint connected to this site.
Zero-downtime deploys on a cheap VPS
Field notes
485 published articles. Essays written here open in place; older tutorials take you back to the site where they first appeared.
Meta's automation capabilities have matured dramatically. Here's what truly automates end-to-end, what still needs human judgment, and where AI agents add real value.
TutorialMaster the principles of natural conversation design: turn-taking, backchanneling, and interruption handling that make voice AI feel human instead of mechanical.
ArticleTCPA, state consent laws, and recording rules form a minefield for AI voice agents. This is not legal advice—consult counsel. But here's what actually applies and how builders navigate it.
CareerYes, and the reason is narrower and better than the usual appeal to rigour. You cannot verify what you do not understand — and verification is now the job. Veracode found 45% of generated code carries a serious flaw. Someone has to be able to see it.
ArticleHow a home services company reduced no-shows by 34% and booking costs by 58% by replacing phone trees with a conversational voice agent. Technical walkthrough.
AI / MLPolished voice agent demos hide systemic gaps. We examine why production deployments struggle with noise, accents, and network conditions that lab tests never encounter.
ArticleMeta's automation capabilities have matured dramatically. Here's what truly automates end-to-end, what still needs human judgment, and where AI agents add real value.
CareerLearn it — it takes about a week and it genuinely helps. Then stop, because every interface improvement is deliberately designed to make your prompting skill worthless, and the vendors are quite open about that being the goal.
ArticleAI isn't replacing ad creative—it's fragmenting it. What works, what doesn't, and how to actually deploy generative creative at scale.
CareerNot as a loyalty exercise. Because the revenue model silently determines which technical tradeoffs are correct — and two teams building identical features under different models should make opposite decisions about caching, uptime and cost.
TutorialMeta Conversions API (CAPI) is how modern ad automation platforms track user actions server-to-server. Here's what you need to know if you're building integrations.
FinTechUnit economics asks one question: does one more customer make you better or worse off? If the answer is worse, growth accelerates the problem — and the fastest way to be wrong about it is to average a cost that isn't evenly distributed.
ArticleMost teams reallocate ad budgets quarterly or monthly. Rule-based automation cuts latency. AI agents cut cost. Here's how they differ.
FinTechYour architecture sets a ceiling on gross margin that no amount of cost-cutting can lift. Software is expected above roughly 75% — and the decisions that decide whether you clear it were made years earlier, usually by someone who never saw the number.
ArticleAI-driven ad automation can scale spend and performance—but one policy violation, bid manipulation, or budget loop breaks trust fast. Here's how to automate safely.
FinTechInfrastructure cost sits in a gap: finance can see it but can't change it, engineering can change it but doesn't see it, and nobody's performance review mentions it. That's an ownership problem wearing a technical costume.
TutorialStop losing revenue to attribution gaps. Learn how to pipe Meta Conversions API data into your CRM for accurate, multi-touch attribution.
FinTechA profit and loss statement is a system with about eight components and a strict order of operations. You already read stack traces. This is easier — and it tells you which of your arguments will land and which are structurally unanswerable.
ArticleGoogle and Meta offer fundamentally different automation models. Performance Max uses structured campaigns with API-driven bidding; Advantage+ relies on opaque algorithmic optimization. Here's what actually differs under the hood.
FinTechThe debt metaphor is better than the people using it realise. Debt has a principal, an interest rate, and a maturity — and the only one of those most teams ever discuss is the principal, which is the least important of the three.
TutorialAutomate creative variant testing across your ad channels with an intelligent agent. Learn how to systematically identify high-performing assets and scale winners at speed.
ArticleHow a Series A DTC brand automated bid management, creative testing, and budget reallocation to cut CAC by 32% in 90 days without sacrificing scale.
FinTechAt a 20% net margin, saving a dollar of cost does the same for profit as earning five dollars of revenue. Engineers are usually the only people in the building who can save dollars unilaterally — and almost nobody frames the work that way.
ArticleGenerative Engine Optimization is real, but not what the marketing claims. Here's what the actual research shows about visibility in AI summaries.
FinTechThere are only four places business value can come from, and knowing which one you're claiming does most of the work. The commonest mistake isn't bad arithmetic — it's claiming a benefit in a category the work doesn't actually touch.
FinTechNone of these look like mistakes. Each is a defensible call that a competent engineer would make — and each puts a ceiling on something commercial that nobody will trace back to a design review eighteen months later.
ArticleRanking on page one doesn't matter if an AI engine already answered the question before anyone clicked through. Here's the structural pattern — question-led headings, direct answers, FAQ schema — that actually wins extraction.
FinTechNot a glossary. Each of these is a rate, a ratio or a stock, they compose in specific ways, and most of the arguments you'll hear in a planning meeting are people disagreeing about a denominator. Including the famous 3:1 rule, which its own author says he explained badly.
FinTechFinOps is usually sold as cost-cutting, which undersells it and explains why engineers ignore it. It's actually a feedback-loop problem: cloud spend is the only significant engineering decision with no signal attached, and everything else follows from that.
FinTechMore than most engineers think and less than the 10x stories claim. The useful version isn't a multiplier — it's four specific channels, each with a formula, and knowing which ones your job actually gives you access to.
ArticleThe search optimization acronym explosion shows how industry terminology evolves faster than practices do. AISO, AIO, and GAIO largely repackage what GEO and AEO already meant.
A working notebook
“Some ideas belong here because they connect directly to projects and services.
Older tutorials still live on Medium, where people first found them. This page keeps the trail intact without pretending everything was published in one place.
How I choose a topic
I would rather publish one useful explanation from lived work than five summaries of things everyone already knows.
Step 01
A topic usually starts as the second or third time the same engineering, product, hiring, or research question lands on my desk.
Step 02
I look for a shipped system, a measured result, a failed approach, or a public artifact that keeps the argument honest.
Step 03
Most technical choices touch something else. The useful links are the ones that help a reader follow that chain without opening fifteen tabs.
Step 04
Some pieces belong on this site; others already have a life elsewhere. I keep the original destination and publishing history intact.
Topics covered
The recurring technical and strategic themes represented across this archive.
A quick orientation
The practical details behind this mixed archive.
Those pieces were published there first. I would rather send you to the original article—with its date, responses, and history—than create a duplicate here.
Thirty-five posts are enough to become noisy. The filters separate original essays from outside tutorials, and permanent page links make it easier to pick up where you left off.
Usually. The essays here tend to connect a technical decision to the project, service, or research context behind it. External posts are often narrower, practical guides.
Please do. Real questions are where the best pieces start. Send the context, what you have already tried, and the part that still feels unclear.
Next step
You do not need a polished brief. A bottleneck, an uncertain architecture decision, or a product idea that refuses to become concrete is enough to start.