The cocktail looked promising. The software bill looked less so. In 2025, Yiru Yao set out to build an iPhone app for the bar cart in her home, where plenty of bottles somehow failed to prevent the familiar nuisance of a recipe demanding two or three more. The idea was modest: photograph what you own, keep an inventory, get recommendations that do not require another shopping trip. The build would take roughly two full days. The education would be considerably more expensive.
Yao used a small parade of contemporary tools: an AI design generator, Figma, a web-app builder, a hosted database, image generation, Google Cloud, Xcode and an assistant that could write code from the terminal. The prototype emerged quickly. Then reality arrived dressed as Apple sign-in, database permissions, a drifting tab bar and a keyboard that refused to behave. Debugging the interface alone consumed more than $20 in AI credits. The full AI bill reached $123.
This is where most launch stories become coy. Yao did the opposite. She itemized the expense, described the shortcuts that poisoned later work and admitted where a visual debugger and screenshots sent to a different assistant proved more useful than the coding agent. Her postmortem reads like a builder talking to builders after the confetti has been swept away.
“Building quickly with AI isn’t cheap, especially when you’re relying on it end-to-end.”Yiru Yao, after building her home bar app
The episode contains the compact version of Yao’s career: a practical irritation, an appetite for unfamiliar tools, and an unusual willingness to record the troublesome middle. Today she is a founding FDE at Glean, the enterprise AI company. The title places her where systems meet customers, where a general platform has to survive a specific organization. Long before the title, she had been rehearsing the same kind of translation.
The sabbatical as workshop
In August 2025, Yao described herself as a software engineer who had recently stepped away from corporate work. She had more than eight years of experience and a problem familiar to experienced technologists: AI was moving quickly enough to make expertise feel oddly temporary. Her response was not to pretend otherwise. She made a curriculum.
The plan had two parts. First, learn the landscape and its foundations. Then choose tools, build projects and document what happened. She worked through introductory courses, prompt engineering, agent architecture and the Model Context Protocol. The inventory she assembled ranged from general chat systems and media generators to coding assistants and agent frameworks. It was broad, but the projects kept it honest.
One experiment began with 57 book notes. Yao had been posting short reactions to Instagram Stories, where they prompted the occasional conversation and then vanished after a day. A local library book club supplied the social spark. She decided to give the notes a permanent room: a digital bookshelf with comments, random selections and a custom assistant able to discuss her collection.
The job sounded tailor-made for AI. It also provided a tidy audit. The source list had 57 items; a reliable tool should return those items, enrich them and stop. One assistant managed ten books after an hour. Another introduced books that were not in the uploaded notes. A third completed the spreadsheet in about twenty minutes, though some image links failed. The custom assistant wandered beyond its knowledge file and behaved differently when the underlying model changed.
Yao did finish the site, in about five hours. Yet the durable result was not the card grid. It was a sharper vocabulary for control. Could a system stay inside a supplied collection? Could a user see what it was doing? Could updates flow through without rebuilding the knowledge file by hand? Her tiny book club had accidentally become a respectable evaluation suite.
She was crossing stacks before AI had one
At Carnegie Mellon, Yao studied electrical and computer engineering from 2012 to 2016. Her project list from those years has the cheerful disorder of someone not yet persuaded that disciplines should remain in their assigned seats. There was a Kinect-controlled quadcopter, built with Python and body tracking. There was WhichCupcake, a social voting site hacked together during a Dropbox summer event. Deep Dreamer applied image filters inspired by Google’s early neural-network visualizations and won the “Coolest App” award at iOS DevCamp in 2015.
Her Devpost profile lists 21 hackathons. That number is less interesting as a trophy than as evidence of repetition. Hackathons force a builder to make an idea legible before it is elegant. The clock punishes abstraction. The audience does not care how promising the architecture looked at 2 a.m.; something must move, answer or light up by morning.
For a 2016 embedded-systems course, Yao and three teammates built Acoustic Phantom. A camera tracked a listener while a pan-and-tilt platform aimed a directional ultrasonic speaker toward that person. A Raspberry Pi handled computer vision, servo control, audio and communication with an iOS app. Hardware, vision and mobile software had to agree about the same room at the same time - a coordination problem wearing the delightful disguise of a speaker that follows you around.
A builder’s map, 2015-2026
The same year, Yao appeared in a place few engineering résumés know how to file: a runway. She joined Katherine Wong and Eunice Oh to design Dauntless for Carnegie Mellon’s student-run Lunar Gala. The collection put delicate floral accents beside structured armor, using foam-sprayed pleather, silk and jersey, with gold-colored wings. The concept was elegance arguing with ferocity and discovering they shared a wardrobe.
It is tempting to treat the fashion show as an eccentric footnote. It makes more sense as part of the main text. Clothes are systems with users, constraints and unforgiving demos. They have to fit a body and permit it to walk. A garment can be visually persuasive on a table and fail the instant a person enters it. Software knows this trick very well.
The useful distance between a demo and a deployment
After university, Yao built a career in software. A former colleague described her first assignment on one team as rewriting an Android project from scratch so the product could launch on a new platform. The same recommendation credits her with owning work across Android, iOS and backend services, as well as testing each release. Her own professional summary emphasizes product-oriented engineering.
That phrase matters. Product-oriented engineers are professionally suspicious of code that succeeds only in isolation. They watch what happens when a feature meets a customer, a permissions model, a slow network or the thumb of someone who did not attend the planning meeting. During her sabbatical, Yao brought that suspicion to the AI market.
Her conclusions were neither worshipful nor dismissive. The tools made ambitious personal projects accessible. They also hallucinated, obscured their progress, behaved inconsistently, generated brittle shortcuts and charged by the attempt. She found ways around some limitations: enrich the data before giving it to the assistant, use a visual debugger for visual bugs, prompt the coding tool to research before patching. In other words, she did not ask whether AI worked. She asked where, under what supervision and at what price.
A side project can be a small thing and still ask a serious question: where does the machine stop being impressive and start being dependable?
Forward-deployed engineering asks the same question in harsher weather. At Glean, the surrounding subject is enterprise knowledge: information scattered across documents, messages, applications and permission boundaries. The public record does not spell out Yao’s individual assignments, and no invention is needed. The title itself explains the shape of the work. “Forward deployed” means close to the place where the technology must become useful.
There is an appealing symmetry in the move. In 2025, Yao deliberately gave up access to the complex systems of corporate work and noted the drawback: fewer natural, large-scale problems. She compensated by curating her own. Books became a grounding problem. A bottle cart became an inventory and recommendation problem. Apple sign-in became a lesson in native constraints. The household supplied a pocket enterprise.
By 2026 she was back inside a company, now working on enterprise AI itself. The curiosity survived the transition. So, one hopes, did the habit of keeping receipts.
A career assembled from contact points
Yao’s public story is not a tale of one grand invention. It is an accumulation of contact points: code touching hardware, an app touching a bar cart, a model touching a book list, engineering touching fabric, a platform touching the actual shape of someone’s work. The lively bit is always the seam.
Her projects also offer a quiet correction to the idea that technical range means collecting tools. She does collect them, energetically. But the memorable details are human. A cocktail recipe should respect the bottles already owned. Reading notes should outlive a social post. A directional speaker should point toward a listener. A dress should let its model walk. The technology changes; the insistence on a use case does not.
That is the useful art in Yao’s curiosity. It is not idle fascination, and it is not the frantic fear of missing a new framework. It is curiosity made answerable to an object, a person and a deadline. Sometimes it produces a prize. Sometimes it produces gold wings. Sometimes it produces a $20 debugging bill and the wisdom to publish it.