Luis Manrique has a diagnosis for an increasingly familiar corporate ailment: the AI demo is charming, the production bill is alarming, and the expensive model learns absolutely nothing about the customer who keeps paying it. A support agent drafts the wrong tone. A salesperson fixes the next action in a CRM. An engineer trims a needlessly elaborate helper. Each correction contains judgment. Yet for most companies, Manrique argues, that judgment vanishes after the request. The meter runs; the memory does not.
Understudy Labs, the San Francisco company he founded with Aamir Poonawalla in 2026, is an attempt to repair that odd arrangement. It sits beside a repeated AI workload, captures traces, turns actual outcomes into evaluations, and lets alternative models compete. A smaller model gets promoted only when it clears a held-out quality bar. The result can run in Understudy’s cloud, the customer’s cloud, or on-premise. The customer keeps the specialist model, along with the prompts, evaluators and routing rules that made it useful.
The immediate pitch is cost. Understudy says it can cut certain frontier-model bills by as much as 80 percent. Manrique has publicly described seeing one repetitive workflow fall from about $400,000 a year to a few hundred dollars a month after a team fine-tuned a smaller open model. But cost is merely the doorman. The larger room contains a more interesting idea: work should produce institutional intelligence, not just an invoice.
“We started Understudy because we believe every company using AI should be accumulating intelligence, not accumulating API bills.”Luis Manrique, Understudy launch
Act I · The educationModels, markets and the middle
That thesis did not arrive in a single flash. Manrique has spent much of his career in the untidy middle where machine learning meets a product, a buyer and an operating budget. He was early at Wildfire Interactive, the social-marketing company acquired by Google. From 2013, he spent five years across Google’s advertising machine-learning and programmatic platforms. The work offered an education in what models look like once they stop being research and start moving money.
In 2017, VideoAmp hired him from Google as vice president of product. He described his strength then as a “pragmatic and collaborative approach” to difficult business and technical problems, and his passion as mentoring product and engineering teams. The language is worth noticing. Manrique does not present technology as a magic trick performed alone. His preferred verbs are social: bring, mentor, collaborate, sell.
At VideoAmp, he worked on infrastructure intended to reconcile the old geography of television with digital advertising. Understudy’s team biography credits him with leading a pivot that took revenue from $5 million to $50 million. He later co-founded Ones and Zeros, a company using transformers for text-to-SQL. It was a first trip through the full founder loop: the technology, the market, and the bracing discovery that good software does not distribute itself.
Instacart brought the strands closer together. As a principal product manager, Manrique worked first on ads machine learning and then on Carrot AI, the company’s first consumer AI team. His name appears on patent applications covering interfaces that group products by similarity, recipe generation and personalization with language models, and advertising systems paced against store inventory. These are not abstract demonstrations. They live among grocery lists, advertiser budgets and the uncooperative details of what is actually on a shelf.
Poonawalla was there too. He spent roughly a decade at Instacart and built advertising serving, machine-learning infrastructure, auction and experimentation systems. Understudy says the pair worked on optimization systems behind business lines that grew beyond $1 billion in annual revenue. One learned how to frame and sell the product; the other learned how to keep the machinery honest. Their reunion would come after an unusually affectionate detour.
Act II · Customer becomes colleagueThe whiteboard interruption
Manrique met Gumloop as a customer. At Instacart, he was trying to put AI tools in the hands of lawyers, finance teams, operators and data scientists. They understood the broken processes around them, but tools built for coders imposed an unnecessary entrance exam. Gumloop offered the expressiveness of a programming language without requiring one. To Manrique, the important discovery was not merely that the software worked. It was that the population of possible builders was much larger than the software industry had assumed.
He became Gumloop’s first major enterprise customer. Later, standing at a whiteboard with co-founder Max Brodeur-Urbas, Manrique began mapping how the young company could sell into enterprises. In his telling, he was midway through one of his “sermons” when Brodeur-Urbas stopped listening and asked a short question: “Why don’t you do it?” Five words are efficient career counseling.
Manrique joined as employee number two, a Member of Technical Staff and, less formally, the “Founding Unc.” He took the job on one condition: he would help Gumloop get far enough to hire a proper sales leader. During his first nine months, the team logged more than 1,100 hours of pitches. Customers arrived in Slack channels by the hundreds and then thousands. Their broken processes traveled straight back to product.
It worked. Gumloop moved into enterprise accounts including Instacart, Shopify and DoorDash. Its public record credits Manrique with closing $1.9 million in annual recurring revenue while leading go-to-market, forward-deployed engineering and data-platform work through a $50 million Series B. The useful part is not the figure alone. Manrique had crossed the same bridge in both directions: user to seller, pain to product, product back to user.
His departure essay, published in May 2026, reads less like a victory lap than a thank-you note. He praised Gumloop’s founders, engineers and customers, then ended with two plain words: “Back to building.” The next problem would combine almost everything that came before it.
Act III · The compounding loopAn understudy earns the stage
The name is theatrical and precise. An understudy watches the lead, learns the role and steps in when ready. In Manrique and Poonawalla’s version, a frontier model begins as the lead. Understudy observes one production workload, builds evaluations from traces and expert review, improves the prompt, tests other routes, and post-trains a specialist if the use case merits it. The candidate does not advance because a leaderboard says it is clever. It advances because it performs the customer’s job.
This insistence on evaluation is the unglamorous hinge. A cheaper route that quietly damages the product is merely an expensive route wearing a discount sticker. Understudy keeps quality, cost and latency on the same screen, promotes only after the candidate passes a held-out bar, and preserves rollback. The system may discover that a frontier model should remain in place. Manrique’s point is not that every grand model is wasteful. It is that a generalist can be comically overqualified for classifying the same message, selecting the same tool or filling the same structured arguments millions of times.
The company’s published examples make the argument in three currencies. On one CRM workload, an Understudy route scored 0.630 against a Sonnet baseline of 0.557 while running at a quarter of the cost. Another operations workload matched performance with an 8-billion-parameter model at 5.2 times lower latency. A sentiment-analysis job labeling 39,962 comments ran at 4.4 times lower full-table cost than Sonnet and 50 times lower than Opus. These are company-reported demonstrations, not universal laws. Their purpose is to show the method: define good, measure it, then let price and speed enter the room.
“I believe intelligence should compound for the company and people that do the work, the same way it always did.”Luis Manrique, “Intelligence Should Compound”
In June 2026, Manrique gave the philosophy its fullest form. The old machine-learning advantage, he wrote, came from owning distinctive data and turning it into a model competitors could not reproduce. Generative AI disrupted that loop. Companies began renting the same capable brain by the token. Their prompts, retrieval and interfaces could still differ, but the intelligence improved for the provider, not necessarily for the customer.
Understudy’s counterproposal is that every correction should accrue. The badly drafted email, the repaired tool call and the engineer’s preference for a minimal helper are not annoyances to be swept from the chat history. They are lessons. Captured carefully, evaluated honestly and used with permission, they can make a narrow system better at the work that produced them. The competitive moat becomes less a static pile of data than a living habit of learning.
Act IV · The social systemRelentlessly helpful
There is a matching human philosophy in Manrique’s public writing. His advice to younger founders is to be “relentlessly helpful”: take the call, try the demo, buy the product and pay it forward. The injunction risks sounding decorative until set beside his own behavior. He bought Gumloop before he joined it. Another founder has credited him with becoming a customer within 48 hours of her company’s pivot. He describes product people as evangelists who should talk to users until the market begins pulling the product from them.
That instinct also explains Understudy’s origin. The company is backed by people from Manrique’s prior chapters, including Instacart co-founder Max Mullen and Gumloop co-founder Max Brodeur-Urbas. Poonawalla brings the infrastructure memory of Instacart. Manrique brings the accumulated conversations. A new venture often looks like a clean beginning from the outside; this one looks more like a network of unfinished discussions becoming a product.
Understudy entered Y Combinator’s Summer 2026 batch as a two-person company. Its public ask is narrow: bring one real production workload where cost or latency has begun to hurt. The modest scope is part of the appeal. One workload supplies traces. Traces become evaluations. Candidates earn promotion. The loop repeats. No model is asked to know the universe before lunch.
Manrique has moved through ads, television measurement, groceries, automation and now inference. The industries change; the fascination does not. He likes systems that improve through use and people who close the distance between a problem and the person able to fix it. Understudy gives that preference a technical form. The expensive model can keep top billing where it changes the outcome. Repetitive work gets a specialist waiting in the wings, listening carefully, preparing to go on.