Machine learning at Coframe, San Francisco
I build the machine learning that decides what you see next.
Recommendations, search and grids for online stores. At home, I answer to two small children.
Up first Bus Stop Confetti
Until you press something, these are my picks. After that, they are yours.
Turn on JavaScript and these toys do real things.
More asks
Ask a figureJump to a drawing
The shelf
This is a real recommender, and it is running in your browser as you read this. Everything on the page belongs somewhere among five topics. When you open or play something, the shelf revises its opinion of you, scores everything again and reorders itself, keeping one place for something you have shown no interest in, on the theory that you may not know yet. It learns only from what you choose to open. Nothing is stored, nothing is sent anywhere, and it forgets you when you close the tab.
Cold start: these are Jonathan’s picks.
01Recommendation rails“You may also like,” done properly.Jonathan’s pick
02SearchWhat the shopper meant, rather than what they typed.Jonathan’s pick
03Feeds & gridsThe same category, in a different order for each visitor.Jonathan’s pick
04The ML factoryWhere a day of clicks becomes tomorrow’s model.Jonathan’s pick
05Head to headThe shoppers decide which model stays.Jonathan’s pick
06The experiment engineEach test, won or lost, suggests the next one.Jonathan’s pick
07“I do it”From “help me” to “I do it.”Jonathan’s pick
08Little boat, fat boatThe little boat is Maggie.Jonathan’s pick
09Anything with wheelsIf it rolls, it gets studied.Jonathan’s pick
10Five books at bedtimeFive is the largest reasonable number.Jonathan’s pick
11Two languagesFive little ducks, in Mandarin.Jonathan’s pick
12Saturday pancakesHe mixes, and she eats the cheese.Jonathan’s pick| Step | On this page | At Coframe |
|---|---|---|
| What gets ranked | Twelve drawings | Live product catalogs |
| What it learns from | The drawings you open and play | Real shopper behavior, from the first visit |
| The model | A Dirichlet over five topics | Learned embeddings and rankers |
| How it stays fresh | Updates after every open | Updates after the next product view and retrains on schedule |
| Where it runs | Your browser, in a fraction of a millisecond | A production ML API, fast enough to add no visible wait |
| How we know it works | You watch it move | Every model ships as an A/B test against what was already there |
Coframe uses AI to improve and personalize online stores. My part is the machine learning that decides what each shopper sees: which products appear beside the one they are looking at, the order of a category page, what search returns and what to suggest to someone hesitating over a cart. I am building it so that a client can ask for something in a sentence and have a working, tested model behind it within hours. Each model checks its data before it trains, retrains every day and replaces the live one only when a fair comparison says it is no worse. The final judge is revenue, measured against whatever the store used before, which has the advantage of being difficult to argue with. The machines below run small versions of the real ideas, and their buttons work.
“You may also like,” done properly.
Every store has a row that says “You may also like.” Ours form a view of the shopper on the first visit and revise it after the next product they look at. A substitute and a complement are different questions, so each gets its own model, and what a shopper has already viewed is something the model weighs rather than a rule it obeys.
What the shopper meant, rather than what they typed.
Search results reordered by what the shopper meant rather than the exact words typed, learned from how people search and what they then buy.
The same category, in a different order for each visitor.
Collection pages and homepages, put in a different order for each visitor. One product view is enough for the grid to reconsider.
Where a day of clicks becomes tomorrow’s model.
Events, features, training, evaluation, publishing and serving, in one place. A model learns the goal of the experiment it serves, which is usually revenue per visit. A client’s request becomes a working model within hours, and the model retrains every day.
The shoppers decide which model stays.
Every model goes live as an experiment against whatever was on the page before, including established vendors in their own slots. A new model replaces the old one only when it is at least as good, and a gain that sits inside the noise does not count. It must also beat the best simple rule, which is more humbling than you might expect.
Each test, won or lost, suggests the next one.
Every test, won or lost, teaches an engine that proposes the next ideas for every client. Agents build the ideas, and people decide which of them deserve a store’s visitors.
Underneath the models are the platform they run on and the method by which all of it gets built.
A client’s request becomes a record rather than a project.
Each client’s request becomes a record that says what to show, what to learn from and how success will be judged, with custom code only where the record cannot say it. A short check proves that the data exists before any training is paid for. Clients are walled off from one another. A new model answers quietly beside the old one before anyone sees it, and the way back is written down before the switch.
The agents build it, and people decide.
I build all of this with a company of AI agents. One plans the work, another audits the plan, and builders take the rows. A grader from a different family of models checks what they produce, on the reasonable theory that nobody should mark their own homework. A piece of work is finished when a script prints PASS, not when an agent says so, and the person who owns a launch decides when it launches.
The same loop runs for every client, every day. Shopper events become features, the features become candidates and rankings, the rankings are served, an experiment says whether they helped, and the answer goes into the next model. A page counts as live only when a new visitor actually sees the model’s choices on it.
Three songs for my family, played over our own photographs.
The songs were made with Suno, and the photographs come from our family album.
Now playing
Follow the stolen ship home.
An action game in four acts, each as long as its song, from “Three Hearts Inside a Stolen Ship” to the lullaby 门前大桥下 on the river home. Run an unwritten road while the Blight eats it behind you; your weapons fire by themselves, and every level is a choice that branches and evolves. With Gentle on, Wren pulls you up whenever you fall, so anyone can reach home.
It plays with a keyboard, a touchscreen or a gamepad, with the sound on, and each act lasts about five minutes.
Game night
Trivia for people who don’t know the answer.
A trivia game for friends who think they know things. Everyone locks in a guess, the guesses become an odds board, and you bet on whoever you think is closest without going over. Play at a live table now, or start a room and text the link so everyone plays on their own phone, with voice chat and AI guessers who explain their terrible reasoning in the group chat.
Play Priced InLet kids meet hard things and find out they can learn them, with someone close by.
Most of what I believe about learning, I believe more since becoming a father.
None of this is ranked. Some things you don’t optimize.
The best sound in parenting is “help me” turning into “I do it.” A balance bike gets there faster than any lecture.
On the walk home along the water, every boat must be assigned a role in the family. The little boat is Maggie. The fat boat, I am told, is Daddy.
Monster trucks, fire trucks, trains and cranes are all examined from as close as an adult will allow, and for as long as one will wait.
Five books, every night, because five is currently the largest reasonable number in the universe. So far, the books are winning.
Our home runs in Mandarin and English. The songs we sang to our kids, they now sing back to us.
He mixes the batter with great seriousness. His little sister supervises and handles the cheese.
A walk home from preschool can take three hours. Every boat, fire truck and unusual machine requires investigation.
Work and home turn out to run on the same method. You watch closely, try something, keep what works and change your mind when the evidence says so. Children do this without being asked. Models have to be taught, and so, now and then, do I.