Working at Ooak Data: Culture, Pace and Who Thrives

The Rhythm They Built

Ooak Data has cut ingestion-to-delivery time by 3x across data from 20 companies while generating revenue "well above 7 figures" from multiple frontier labs, all with a five-person team split between Paris and San Francisco per Y Combinator's directory. The question isn't whether a three-person founding team can run a revenue-generating AI lab from two continents. It's whether the operating rhythm they've built survives the hiring wave they've just triggered.

The company sits at an unusual intersection: a Y Combinator Summer 2026 company with a Paris headquarters at Station F, a San Francisco presence, and a product that turns live enterprise data into reinforcement learning environments for frontier labs. Founded in 2024 by Pierre-Louis Vouteau, Grégoire L., and Thomas Aubry — three operators who, per their YC profile, "spent our careers inside the messy company systems we now replicate." Ooak's lean structure and hiring patterns signal both its operational priorities and the kind of candidate it seeks.

Authority rests with the three cofounders. No separate C-suite exists yet. The first-party board data shows 12 salaried roles posted across Paris and San Francisco. Paris carries the core research and engineering weight; San Francisco handles lab relationships and revenue. The Chief of Staff and Head of Operations roles, both based in Paris, signal that the founders are offloading operational execution to keep decision cycles tight.

Pace is set by the product cycle. The team hosts hackathons (one in May 2026 drew 11 data teams to the Paris office) not for culture points but to stress-test the platform on complex permission and multi-domain setups. The work is outward-facing: the product only matters if agents perform better on real workflows than on synthetic benchmarks.

Decision-making flows from the problem space. The founders built the anonymization pipeline and the task calibration framework themselves. New hires, particularly the Founding ML Engineer and the "bras droit fondateurs" (chief of staff equivalent) roles advertised in November 2025, inherit defined technical boundaries: extend the RL environments, scale the data partnerships, harden the security posture. The Head of Engineering role suggests the codebase has outgrown founder-only maintenance. The Head of Operations and Chief of Staff roles suggest the logistics of sourcing data from 20-plus companies, managing anonymization compliance, and coordinating with frontier labs now require dedicated ownership.

The dual-location structure adds friction by design. Paris is where the data lands, the models train, and the environments evolve. San Francisco is where the labs sit and the contracts close. The GTM and Account Executive roles in San Francisco report into a founding team that operates primarily from Station F.

This structure holds while the team is small and the product is the only priority. The open question — the one the next six months will answer: whether the same rhythm survives when the headcount grows, the Paris office fills, and the San Francisco outpost needs its own operating cadence.

The Operating Principles That Rebut the Paradigm

Ooak Data's operating principles read like a rebuttal to the dominant paradigm in AI evaluation. The company was founded on a single, specific conviction: synthetic benchmarks cannot capture the messiness of real enterprise work, and any evaluation framework that relies on them produces misleading signal. That belief shapes every technical and product decision the lab makes.

The principle shows up first in the data itself. Ooak does not generate training environments; it sources them. The company connects directly to customer tooling and builds what it calls digital twins: structurally identical copies of an organization's workflow graph, stripped of identifying information through an automated multimodal anonymization pipeline. Names, dates, and proprietary content are transformed, but the org chart, permission boundaries, and cross-tool dependencies remain intact. The research team describes this as "multimodal from the start": documents, conversations, project management artifacts, and organizational structure captured together, not as text with other modalities bolted on later. The distinction matters because real work moves across modalities; an agent that can summarize a document but cannot navigate the approval chain embedded in Slack and Jira will fail in production.

The second principle: evaluation must target the frontier, not the average. Ooak's tasks are calibrated on the latest frontier models and designed to expose weaknesses, not confirm strengths. As models improve, the environments evolve. "You are always testing at the edge of capability," the company states. This is a moving-target philosophy: the benchmark is not a static dataset but a living suite that advances with the models it measures. It also explains why Ooak positions its product for frontier AI labs, enterprise AI teams, and AI startups simultaneously: each segment needs to know whether an agent can complete a multi-step, multi-tool workflow under realistic constraints, not whether it can answer a single-turn question.

A third principle emerges from the anonymization architecture: privacy and utility are not a trade-off but a joint engineering problem. The pipeline must preserve the structural complexity that makes the environment realistic while guaranteeing that no sensitive data leaks. The company's own messaging frames this as the "hard part": sourcing real data and making it safe. The same constraint appears in the broader industry conversation around agent security, where the tension between locking data down and enabling agent access is described as the central blocker to deployment. Ooak's approach, structured, permission-aware digital twins, is effectively a technical answer to that industry-wide dilemma.

The fourth principle is organizational: the lab operates as an applied research outfit, not a conventional SaaS vendor. The revenue figure, "well above 7 figures," and the named YC partner Diana Hu signal that the commercial traction is real, but the self-description as an "applied AI research lab" and the stated ambition to launch a dedicated AI data research lab by 2027 (the Alexandria project) indicate that the founder's time horizon extends beyond the current product cycle. Alexandria is described as "the world's largest library of real-world business workflow datasets": a public good ambition layered atop the commercial engine.

These principles cohere into a recognizable culture: technical rigor over benchmark theater, realism over convenience, and a willingness to solve the unglamorous data-plumbing problems that block agent deployment. The company does not hide the difficulty. Its own copy admits that real org charts, permissions, and cross-tool dependencies "don't come out of a generator" and that sourcing and sanitizing them is "the hard part, and it's what we do." That framing — owning the grind is itself an operating principle, and it filters for a specific kind of candidate.

The hiring signals read like a filter for that candidate: someone who has already wrestled with the gap between benchmark scores and production reality. The careers page states it directly: "If you care about real-world AI performance, not just benchmark scores, we want to talk to you." That sentence does more framing work than any competency matrix. It tells you the team optimizes for practitioners who have felt the pain of agents failing on multi-tool, permission-heavy workflows inside actual organizations, not researchers chasing leaderboard positions on static datasets.

The evidence sits in the roles themselves. Every open position carries a seniority floor of six years, except the data engineer role at three-plus. The company, at five people total per Y Combinator's directory, cannot afford passengers. Each addition must expand the surface area of what the team can ship.

The founders' backgrounds reinforce the pattern. Thomas Aubry, founder CTO, led the Data & AI team at PayLead and worked as an applied ML scientist at Samsung AI. Pierre-Louis Vouteau, cofounder and CEO, was COO at Visely, a UK ed-tech company, with prior stints at Gopuff and L'Oréal. The third cofounder, Grégoire, is referenced in the YC profile but not detailed publicly. All three, per the company's own description, have built their careers on the messy company systems they now replicate. That phrasing is deliberate. They are not looking for people who have read about enterprise data plumbing; they want people who have unclogged it.

Two stated operating principles function as hiring filters. "Small team, high trust: We hire people who take ownership. No micromanagement, no unnecessary process. You will have real impact from day one." And: "Intellectual honesty: We say what we don't know. We question our assumptions. We'd rather be right slowly than wrong quickly." In practice, these translate to a bar that screens for autonomy and calibration. A candidate who needs detailed tickets, weekly syncs, or permission to refactor a pipeline will struggle. So will one who presents synthetic benchmark improvements as evidence of progress. The work of sourcing real enterprise data, anonymizing it into digital twins that preserve org charts and cross-tool dependencies, then building RL environments calibrated against frontier models rewards engineers who can work through ambiguity without losing rigor.

The technical stack implied by the product reinforces the same signal. Multimodal ingestion (documents, conversations, project management tools, org charts), automated anonymization pipelines that preserve structure while stripping PII, and RL environment generation that evolves alongside model capabilities: this is not a standard MLOps loop. It requires fluency across data engineering, privacy-preserving transformation, and agent evaluation methodology. The "applied AI research lab" framing on the website is not marketing; it describes a workflow where research and production are the same activity. "We publish what we learn, but everything we build goes to production"; that line from the careers page collapses the usual research-engineering handoff. Candidates who have only operated on one side of that boundary will show friction.

The GTM and partnership roles carry the same DNA. The account executive for data partnerships and the GTM role targeting AI labs both require selling a product that does not yet have a category. The buyers are frontier labs and enterprise AI teams who already know synthetic benchmarks are insufficient. The sale is technical, consultative, and grounded in the specific failure modes of current agents: multi-step workflow breakdowns, tool-use errors, permission handling. A candidate who has only sold API access or seat licenses will not have the vocabulary.

There is no public interview rubric, no leaked packet, no Glassdoor trail deep enough to reconstruct a stage-by-stage process. The company is too new, too small. But the aggregate signal is consistent: Ooak hires people who have already done the hard version of the work — real data, real constraints, real deployment and want to do it again at a higher level of impact. The bar is not a puzzle or a whiteboard exercise. It is a resume that shows you have lived in the problem space they are now productizing.

The Silence Is the Signal

The shortest honest answer: almost nothing on the record. Ooak Data employed five people per Y Combinator's directory, three cofounders and two early hires, and the company graduated from Y Combinator's Summer 2026 batch only months ago. Glassdoor lists five reviews total, all anonymous. No former or current employee has spoken publicly by name about daily life inside the company. That vacuum is itself a signal: at this stage, "culture" is whatever the founders decide it is on a given Tuesday, and the first hires will write it in real time.

What exists are proxies. A September 2026 LinkedIn post from Grégoire L., a cofounder given the recruiting context, framed the pitch bluntly: "The real pitch: you'd be building alongside Thomas Aubry, our CTO. He's a crack. If you want the kind of year where you learn faster than you thought possible, that's the reason to come." The same post noted "Contracts signed, revenue in, and basically everything still to build." That phrasing — "basically everything still to build" is the clearest cultural marker available. It signals a team that treats incompleteness as a feature, not a bug, and expects hires to operate without guardrails.

The Glassdoor reviews, while anonymous, are too few to generalize from. With a headcount of five per YC's directory, any specific detail would de-anonymize the writer instantly. Treat the sentiment as directional, not evidentiary.

No departure announcements, public retrospectives, or on-record critiques exist. The company's Y Combinator page lists Diana Hu as primary partner and shows the team as "Active" with status unchanged since the batch. That stability, with no visible churn in the first months, is the only verifiable retention data point. But with five people, one departure would be a 20 percent turnover event. The sample is too small to generalize.

The hiring posts themselves reveal how the company wants to be perceived. The same LinkedIn post advertised "CDI or freelance, both work for us, as long as you are in Paris (9e)": a flexibility unusual for early-stage AI labs, which typically demand full-time, in-office commitment. The roles listed (Lead Data Engineer, Senior SWE, Senior Data Engineer, ML Engineer, Chief of Staff, Founding Marketing Lead) carry equity bands of 0.10–1.00 percent and Paris-market salaries (€55K–€100K base). That compensation structure, published transparently on the YC jobs board, suggests a team trying to attract senior talent without overpromising, or at least without hiding the numbers.

What candidates should take from the silence: you will not find a culture document, a values deck, or a Glassdoor trail deep enough to pattern-match against. You will find two technical founders (Aubry, formerly Head of Data at PayLead and that role; Vouteau, formerly COO at Visely, Gopuff, and L'Oréal) and a third cofounder, Grégoire, whose background the public record does not detail. LinkedIn shows 12 employees with several named beyond the founders. If you join, you become part of a team still defining its culture. The culture is what you and they negotiate in the first ninety days.

Who Thrives, Who Doesn't

The company's own career page distills its operating DNA into four lines: real data over synthetic shortcuts, research that ships, small team high trust, intellectual honesty. Those aren't slogans — they're filters. A candidate who needs a spec before writing code, who optimizes for benchmark scores on static datasets, or who expects a manager to break down tickets will not last long. The work is the opposite: you ingest messy, multimodal enterprise data, build anonymization pipelines that preserve organizational structure while stripping PII, and ship RL environments that frontier labs use to stress-test agents on multi-step, multi-tool workflows. The cofounders, Pierre-Louis, Grégoire, and Thomas, know that terrain firsthand. They hire people who can operate at that same level of ambiguity.

The Profile That Succeeds

Ownership without supervision. The careers page states it plainly: they hire people who take ownership, with no micromanagement or unnecessary process, offering real impact from day one. With roughly 12 salaried roles on the board, every hire is a force multiplier. The founding ML engineer role and the chief of staff role (chief of staff) are explicitly framed as founder-adjacent: you're not executing a roadmap; you're helping define it.

Comfort with real-world data grime. The company's thesis is that synthetic benchmarks test theory; real org charts, permission graphs, and cross-tool dependencies test practice. Candidates who have only worked on clean academic datasets or public benchmarks (MMLU, GSM8K, SWE-bench) often underestimate the engineering surface area: multimodal ingestion, automated anonymization that preserves relational structure, task calibration against frontier models that improve weekly. The team has done so at that rate: that speed comes from engineers who treat data plumbing as a first-class product problem, not a preprocessing afterthought.

Intellectual honesty as a daily habit. As the careers page states, they say what they don't know, question their assumptions, and prefer being right slowly over wrong quickly. In a domain where the evaluation target (frontier agent capability) moves monthly, this isn't philosophical — it's survival. The RL environments are built on that principle. An engineer who ships a benchmark that flatters the model is doing the opposite of their job.

Pragmatic research mindset. This philosophy filters out pure researchers who want publication counts and pure engineers who treat evaluation as a checkbox. The sweet spot is someone who can design a rigorous ablation study on Friday and ship the resulting environment to a frontier lab on Monday.

Low ego, high trust. The values list "Collective, team first, low ego" and "Kindness, at the heart of every interaction." In a 12-person team working with multiple frontier labs (named customers are confidential, but "world's leading AI labs" is the claim), reputation compounds fast. A brilliant jerk destroys more value than they create because the product is trust: labs trust Ooak's environments to be structurally faithful, companies trust the anonymization pipeline with their data. One leak or one hallucinated benchmark breaks the business.

The Profile That Struggles

Process-dependent operators. If you need Jira tickets groomed, design docs approved, or a product manager to prioritize your sprint, you'll wait forever. The "small team, high trust" model means the person closest to the problem decides. The chief of staff role is described as helping on "strategic and operational challenges an ambitious AI startup faces": not running standups.

Benchmark chasers. Candidates whose mental model of progress is "beat the leaderboard on Dataset X" will misread the mission. Ooak builds the next dataset: the one that breaks the model that just beat Dataset X. The work is adversarial by design: "Multi-step, multi-tool workflows built on that principle."

Synthetic-data purists. The company's founding insight — that these dependencies cannot be generated — is a direct rejection of the synthetic-data scaling hypothesis for agent evaluation. Researchers who believe enough compute and a good generator solve the data problem will fight the core premise every week.

Remote-first expectations without Paris/SF presence. Headquarters is 60 rue François 1er, Paris 75008. The board lists roles in Paris (Head of Engineering, Head of Operations, Chief of Staff, ML Engineer) and San Francisco (GTM AI Labs, Account Executive - Data Partnerships). The team is small enough that physical co-location matters for the high-bandwidth, low-process collaboration the culture demands.

Short-horizon optimizers. The six-month plan: acquire and anonymize hundreds of data ecosystems from real-world companies to become the leader of real business workflow data. Fundraising kicks off November 2025. The timeline is aggressive, the technical bar is rising (frontier models improve, so environments must get harder), and the revenue is already seven figures. The ambition value — "we commit, we move fast, with tenacity and efficiency" — means the pace is set by the market, not by a sustainable-velocity framework.

Ask yourself: Have I shipped a data product that real customers paid for, where the data was messy, the schema was implicit, and the evaluation was "does the agent actually complete the workflow?" If yes, and you can articulate where you'd rather be right than fast, Ooak is a rare fit. If your best work happened on clean benchmarks with clear metrics and a manager who shielded you from ambiguity, the friction will compound daily. The company doesn't hide this — the careers page, the YC profile, the LinkedIn posts all describe the same hard problem. They're not screening for talent; they're screening for alignment with a specific kind of difficulty.

The rhythm they built — 3x faster ingestion, 20 companies' data flowing, revenue climbing — was forged by three people who knew the mess from the inside.

Working in AI? Zero G Talent tracks the openings: see every open Ooak Data role, browse AI jobs, the companies hiring, and the people building the field.