Skip to content

Insights

Notes from the studio.

Building in public: what I'm making, what broke, and what I'd do differently.

Trained to listen: where my engineering judgment comes from

The nine posts before this one are about code. This one is about where the judgment underneath them came from: a house of two immigrant engineers, junior tennis, recovery in 2014, a decade of intake, assessment and crisis work, and what all of it does to the way I build.

Shipping a native iOS app as a web studio

Four of my apps are on the App Store, and every one of them compiled long before it shipped. SwiftUI is familiar for about a day, then six things stop working the way CSS taught you. The pipeline that gets a signed build onto somebody else's phone is the part nobody warns you about, including the rejection I could not fix in code.

The old browser you forgot to test

One word on line 18 of my stylesheet decides whether the header stays pinned on an iPhone. The iOS Safari position fixed bug, the feature query that gates the fix honestly, and the measurement that found a rule on my own index assuming a browser twenty-six Chrome versions newer than my written floor, which I fixed the same day.

Video on the web that plays everywhere

A video that plays on every machine you own and shows only a poster on somebody else's older iPad has at least five unrelated causes, and the browser will not tell you which one you have. Here is the ladder I walk, cheapest rung first, and the two-line ffmpeg recipe that settles most of the argument before it starts.

The loading experience is the product

A scrub film on my products page is 8,310,726 bytes. Its poster frame is 42,046 bytes, and the poster is the only one of the two most visitors ever see. What the research on skeletons and spinners actually says, what the Core Web Vitals actually measure, and the layout shift my own reveal animation was covering for.

Grace Tarnasky's site, from launch-ready to public

Her palette arrived as a document she had written herself, the headline on the home page is her own sentence, and the fork carried three themes before the quiet one won. Here is what the site is made of, and what ten gated days bought before anybody outside her family could see it.

The tool bench a one-person studio builds for itself

A private bench of ten apps, no accounts, no servers, nobody to support. Here is what is on it, how the slide proofer works out where a platform will crop before the file leaves my hands, and what happened when I ran a slide I had already posted through my own guides.

A booking engine inside the CMS: slots, guards, clocks

Two visitors reach for the same half past ten on a Friday, a second apart. No lock and no unique index can settle it, because the appointment does not live in my database at all. Here is what settles it instead, and the window that stays open behind the answer.

Wireframes that survive contact with the build

Most wireframes get approved and then quietly lose the argument to the browser, because they encoded no constraint the build could not ignore. Here is what mine have to carry: the empty state, the error state, the too-long state, the width the layout really changes at, and a short list of things I refuse to decide.

The fine print on AI video in 2026: what the terms say

I spent a day reading six AI video licences on the vendors' own pages instead of trusting the model cards. One of them forbids commercial use of the output without written permission. One of them excludes the United States from its territory. And the watermark you can switch off is not the one that matters.

A native investing app in Rust that vetoes its own orders

For the first week of this project's life, an account that came back empty read as approved. Every percentage rule in the risk engine was guarded against dividing by zero, and a guard like that carries a second meaning nobody intends. The trading terminal I open-sourced in June is one long argument that a check which cannot run has to answer no.

Color for screens that aren't yours

I ran my own studio palette through the WCAG 2.2 contrast formula as a set instead of one pair at a time, and found two shipping failures, including the link color in every post on this site. The math, the fix in OkLCh that shipped the same morning, the one ground still short, and the tests that catch it next time.

Custom mental health web apps whose answers never leave the device

What a practice gets when it buys one of these is a tool that holds nothing. No row, no account, no export, and therefore no retention schedule to write, no breach plan to rehearse and no deletion request to route through a person three years from now. That is a storage decision, made before any screen is designed, and it decides most of what comes after it. It also has a limit worth being precise about, because the page a tool sits on is a separate question from the tool.

MiniMax H3: the Hailuo image-to-video model, in practice

H3 sits third on the image-to-video arena, generates its own stereo audio in the same pass, and turns a client photograph into five seconds of video for forty cents. It also ships open weights under a licence that excludes the United States by name. Both of those are true, and a studio has to hold both.

What a therapy practice website should never ask for

The federal definition of individually identifiable health information covers anything a provider receives that relates to a person's mental health or condition and can be tied back to them. Not diagnosis, not treatment: condition. So the website for my own counseling practice never asked for a history, never held a calendar and never took a card. Every other decision on it followed from that one word.

A status file per bot: how a small office of agents reports in

I run a small office of bots across two runtimes, and I can see every one of them at once: no message bus, no database, one small JSON file each. A dashboard reads that folder every ten seconds and draws the studio as a floor of desks. Here is the whole mechanism, including the freshness rules and the route I had to close on the day I opened it.

StayEugene: when one house is five bookable listings

StayEugene sells five short-term rentals direct, and four of them are the same building cut at different sizes. That makes availability a question about a set of listings rather than a row in a table, and it is the part of a booking site that a marketplace was quietly doing for you all along.

Which AI video model should a small studio use in 2026

OpenAI removes Sora from its API on 24 September. I re-priced the whole field off vendor rate cards rather than blog posts: what five seconds of video costs on each model, where the arena actually ranks them, and the licence clauses that decide it before the quality does.

World Labs Atlas: what a world model means for AI and physics

World Labs announced Atlas on 1 September 2026, a model trained on 3D Gaussian splats alongside text, images and video. Who built it, what a splat actually is, what the physics benchmarks say, and what a small studio can do with the parts that already ship.

A counselor directory on iOS: what moved to the phone

The boundary between a web directory and its iPhone app is a decision about what deserves to travel. Mine let 589 verified providers cross, in four public JSON files, with the crisis numbers compiled into the binary so they survive with no signal. It left behind the unclaimed license records, the ranking, the accounts, the analytics, and any reason to ask where you are.

Designing without slop: the rules I gave my AI

Every AI design tool lands in the same place: purple gradients, Inter set large, a centered hero over three feature cards. My answer was not to complain about it but to write a law an agent can be held to. Here is that law, with the real values behind it.

One button, no reading, no login: designing panic.btn

There is no account to make, no menu to open and no paragraph to read. Press the one shape on the screen and it becomes a breathing pacer, vibrating on each phase change so the pace can be followed with your eyes shut. This is what that press does, where a counselor's vocabulary shows up in the interface, and where the app stops.

iLovable: a box of therapy tools that stays on the phone

A parts map, a CBT chain, a mood check-in and a day count, in an app that downloads at 13.5 MB and sends nothing to me. The exercises came off paper I handed people in a counseling room for ten years. Most of the design work was deciding what the screen would refuse to do.

What on-device AI costs: a voice journal at 813.4 MB

Mutterly runs on an iPhone XS, a phone from 2018, because its language model rides inside the download. Apple's own on-device model ships almost nothing extra and starts at the iPhone 15 Pro. Five model years of phones sit between those two floors, and that gap is the whole reason for the size.

Twelve neighborhood guides and one valuation form

A market report is accurate and generic; a neighborhood guide is specific and goes stale. Grace Tarnasky's site runs both and joins them with a five-digit number. What each half is for, where the data comes from, what a guide is not allowed to say, and why the valuation form refuses to guess.

What a licensed psilocybin practice needs from its site

A regulated practice cannot buy a single click, so organic search has to carry everything. Here is what that changed: an age gate that keeps the page crawlable, booking built into the CMS with the acknowledgments recorded, mail that sends through one provider and receives through another, and a review card I shipped and then switched off for good.

The scroll-scrub stack: what actually makes it smooth

A scrubbed film either reads as liquid or as a slideshow being dragged, and I finally measured where the difference lives. It is the encode. One clip, three encodes, sixty seeks each: a median of 3 ms against 15 ms against 32 ms, with the whole frame budget sitting at 16.7. Here is the loop, the encode, the dead zones, and the one line that makes iPhones cooperate.

Mobile-first, one thumb

Mobile first design, as I actually build it: finger-sized targets, Pointer Events instead of mouse events, and nothing important hidden behind a hover. Then the morning my own audit failed a component that was the right size all along, because the test drove a narrow desktop window and called it a phone.

One engine, three client sites: the fork per client model

Three live sites share one CMS engine and share nothing at runtime: separate repositories, separate keys, separate databases, separate hosting accounts. The bill for that isolation arrives every time I fix something in the engine, because a fix lands in one copy and has to be carried to the rest. Here is the ritual, the path a real security fix took, and the four ways the sweep has quietly failed on me.

Finding a counselor in Oregon

A person looking for a therapist has one word for the problem, the name of a town, and about a minute of patience. Here is what ORCounselors.com does with that: one search box, a five-minute wizard, four access filters, and a ranking that sorts by verification first and locality second, until somebody asks to be seen in person.

IDX by Ulric: what the three tiers actually include

The product page prints three prices, $299, $599 and $999 a month, and the sentences around them were the slow part. Here is what each tier includes, what the three pinned scroll stages are demonstrating, and how a license agreement ends up writing your marketing copy for you.

What $99 a month buys a clinical practice

Two tiers on a page nobody has to call about: Foundation at $99 a month, Foundation plus Local at $199, setup $500 once. What each one includes, why the Stripe hand-off happens after the intake instead of from a buy button, and the three rules that decide how a therapy practice site gets built.

A social composer that exports in the browser

The Ulric Composer designs Instagram posts and reels and renders the finished file in the page: a real PNG off a 2D canvas, a reel recorded frame by frame, no server round trip. How the export works, what WebKit refuses to paint, and the one number the browser and ffmpeg both have to agree on.

Three small iOS apps, and what each of them taught me

Three small iOS apps, one of them 340 times the size of the smallest, and all three declare the same thing on the store: data not collected. The spread between those numbers is the price of keeping a promise on the device, and it is the most useful thing I took out of building them.

Two kids built games, so I gave games a content type

Lulu is 10, Otto is 11, and they are my niece and nephew. Over a handful of afternoons this summer they finished three browser games they wanted to play, and I had to decide where those games would live on my site. Three hand-written pages would have rotted by autumn, so games became a content type instead: a table, a folder scan, and a page that ends with what the game taught the kid who made it.

Procedural geometry vs Blender MCP: where I drew the line

I rebuilt the procedural world demo to settle an argument with myself: draw the geometry in code, or model it in Blender now that an agent can drive Blender over MCP. Four worlds of fences, trees and desks cost 19 to 53 draw calls a frame in code. One running dog cost a 0.91 MB file and three trips back to the prompt.

Anatomy of an agent harness: hooks, loops, memory, RAG

The model is the one part of an agentic system I did not build and cannot pin down. Everything around it is scaffolding I can test: hooks that block, a loop that verifies, files that remember, a ladder that routes, retrieval that has to stay true. Six components, what each one decides, and the failure that put it there.

An IDX site on a RESO Web API feed: what the docs leave out

The RESO Web API standard guarantees a small floor: paged results, an explicit field list, and a link to the next page to follow, and nothing more. Everything above that floor is your MLS's choice, and almost none of it is written down. Build notes from a feed integration that went live in August 2026.

28,000 listings on one map with MapLibre and no API bill

Every active listing in the state lands on one map, drawn by MapLibre GL JS on ordinary shared hosting, with nothing metering the page loads. One version-stamped GeoJSON off the sync cron feeds the pins, and a self-hosted PMTiles archive is measured and staged to answer for the basemap the free hosted style is drawing today. Measured numbers, the compliance furniture, and the version pin that a release note forced.

When an image pipeline fails silently: a disk-quota postmortem

A listing page said "48 photos" above a grid of blank tiles, and every layer underneath it, from encoder to cache to queue, had reported success. The account had hit its disk quota, the image encoder wrote nothing and said it had worked, and the cache filed the empty result as truth. Here is the whole chain, and the four changes that shipped the same morning.

LLM model routing: a ladder, a cascade, and a kill switch

LLM model routing is two problems: a ladder that picks the model tier by what a wrong answer costs, and a cascade that answers when a provider dies. Here is how both are built, with the real chain, the real prices, and the paper that explains why the second one will not save you money.

Agentic loops in production: generate, deploy, verify, repeat

Seven weeks after saying loops should run the business, I have two shapes going: one built on a five-minute clock, one that checks in every Monday. The weekly one reads Search Console; here is what the summer taught, including the guardrail that spent two weeks announcing a login prompt as a finished report.

Graph RAG vs vector RAG: when the graph earns its keep

Vector retrieval answers the question of what resembles this text, while a graph answers what is connected to it, and the two fail on different questions. The benchmarks say the difference is large on relational questions and close to nothing on lookups, which makes this an accounting decision rather than an architecture debate. What each one costs, the hybrid I actually run, and how to decide before you buy an index.

RAG that stays current: knowledge that ships with the feature

Most retrieval systems do not fail at retrieval: they fail because the paragraph they found describes software that no longer exists. Keeping RAG up to date is a scheduling problem before it is an engineering one, so I made it a rule: an admin feature is not done until the migration that ships it has also written the chunks that describe it, and the embed step has run.

Grok Bot, and the question every small business now has to answer

xAI shipped agents that get their own cloud computer and sign into your tools like you do. My feed called it god mode. The interesting part is not the army of bots. It is that renting one and owning one are different bets, and most businesses only need one of them.

Claude Code hooks: the guardrails that never ask the model

A model can be talked out of things: worn down by a long context, swayed by a file it read, or simply wrong at one in the morning. A hook cannot: it is code that runs at a fixed point in the agent loop and returns a decision the model gets no vote on. The hook lifecycle, the exit code that actually blocks, and the two guards I run on every session.

Setting up an AI agent for a realtor who was out of hours

A realtor I work with was never short on leads. She was short on hours. Instead of another dashboard, she got an agent harness: a folder of files on her own Mac that reads everything, drafts carefully, and publishes nothing.

Most of my AI bill was re-reading my own transcript

My usage dashboard blamed subagents and had a fix ready. Twenty minutes of tallying the raw transcripts said the opposite: the ten most expensive sessions of the week spawned no agents at all, and almost every token I paid for was the same conversation being read again.

The SEO skill that has to show its receipts

Most SEO advice written for AI agents is a checklist, which is how you get a confident audit of a site nobody looked at. So the studio built one that cannot make a claim without a fetched artifact behind it, then pointed it at five sites in a day.

The harness reads my texts now

I text myself X posts about agent workflows and never read them twice. So the studio harness got read-only access to the thread, and the pile became three shipped upgrades: a blind duel for risky answers, a knowledge graph under the assistant, and a rulebook half its old weight.

Building a real estate site for a realtor in The Dalles

Grace Tarnasky messaged me on a Thursday and had a working site on Saturday. A walkthrough and a look at the decisions: why her listings are real pages instead of an IDX iframe, why six neighbourhood guides matter more than a search box, and what a fork of an engine buys you.

How a real estate agent gets recommended by an AI assistant

Buyers have started asking ChatGPT to recommend an agent. That answer is assembled from sources a machine can verify, which is a different game from ranking, and a more winnable one.

Why your therapy practice does not show up on Google

I ran a private practice and built the directory 16,800 Oregon providers sit in. Here is the honest reason most counseling websites are invisible, and the order I would fix it in.

The questions similarity search can't answer

The assistant on this site can find anything that resembles your question. It cannot yet follow a chain of facts spread across three notes. Why that failure is structural, what a knowledge graph actually is, and how I plan to wire one into the studio's retrieval without tearing out what works.

Scroll-scrubbed video, the cheapest premium feel on the web

I taught the studio toolchain the Apple-style scroll-scrub technique, then built eight live demos from five-second phone clips to prove it. What makes a scrubbed page feel premium, the plumbing nobody mentions, and where the technique actually earns its keep.

AI slop, and the files that keep it out of my projects

Every AI project drifts toward the same fonts, the same gradients, the same sentences. Mine do not, because every agent I run inherits a small set of law files. Here is the whole playbook, structure and all.

The agency that never sleeps: running a business on AI loops

Loop engineering is taught for building software. The bigger idea is using the same loop to run the business: point an agent at a number, give it a scoreboard it can check itself, and let it improve your SEO, ads, and AI visibility for months while you sleep.

Real estate runs on too many dashboards

Every realtor I've worked with pays for five systems that don't talk to each other. The fix isn't another integration. It's one system that owns the whole flow, from listing to lead to closed.

The baton passes

Why a new software studio in Eugene, Oregon is named after a man who spent forty years moving packets.

studio
story

Learning to scale SEO and build organic traffic

Top-3 rankings for the searches that matter didn't come from tricks. They came from structure, patience, and publishing things worth finding. Here's the playbook I had to learn the slow way.

Building trust with thousands of providers in Oregon

A directory is easy. Sixteen thousand indexed providers is a spreadsheet. The hard part, and the real product, is trust, and it gets built one conversation at a time.