Before Maayan Salom knew the phrase “data observability,” she knew the feeling that gives the category its reason to exist. She was in eighth grade, playing Football Manager, maintaining spreadsheets of players, attributes, and statistics with the concentration of a small club’s recruitment department. The numbers promised order. The match results occasionally declined to cooperate. Something, she felt, was hidden. Something did not add up.
The game did, in fact, contain hidden metrics. Salom later admitted that the concealment probably made it more entertaining. Professional life would offer a less charming version of the same plot: a dashboard that looks authoritative, a pipeline with a quiet defect, a report whose confidence exceeds its accuracy. Her recurring question became both instinct and occupation: “This can’t be right.”
Today Salom is co-founder and CEO of Elementary, a company that builds data observability and reliability software. The description sounds orderly. The problem is anything but. Modern companies move information through warehouses, transformations, dashboards, machine-learning workloads, and now AI agents. Each handoff offers another opportunity for freshness to fade, definitions to drift, or a failure to disappear beneath dozens of louder but less important alerts.
“People trust data to be true.”Maayan Salom
That short sentence carries the weight of Salom’s work. A wrong number has a measurable cost, but mistrust compounds differently. Once people have been burned, they keep private spreadsheets, rerun queries, ask colleagues to confirm the obvious, or quietly stop consulting the official dashboard. The expensive data stack remains. Belief leaves the building.
Reliability, imported from another department
Salom arrived at the problem through security. Her professional record begins with years in Israeli Military Intelligence’s Unit 8200, moving through data analysis, cyber analysis, security research, and team leadership. She later worked in incident response at Team8 and Sygnia, where her roles included technical team lead and incident response manager. The vocabulary changed from role to role, but the operating rhythm stayed familiar: establish visibility, notice anomalies early, understand the blast radius, and respond before damage travels.
Data teams had similar stakes without equivalent habits. Software engineering had monitoring and mature operational practices. Security had detection and response. Data work, Salom and her future co-founder Or Avidov observed, often relied on checks assembled after trouble had already become visible to a customer or executive.
The founding partnership began socially. Salom and Avidov were longtime friends who met through Avidov’s wife, who had worked with Salom. Their company handbook confesses that too many double dates were derailed by conversations about startups, data, and engineering. Eventually the conversation escaped dinner and acquired a cap table.
In August 2021, they began researching data observability. They did not hide behind a survey link. In a few weeks, they spoke with roughly 70 data teams, drawing from colleagues, friends of friends, and a substantial amount of cold outreach. The questions were plain: Do you trust your data? Can you see problems? How do you detect them? Which solutions have you tried, considered, or rejected?
A pattern emerged. Existing tools could be complex, rigid, or foreign to the way analytics engineers already worked. The founders chose a demanding constraint: Elementary should feel native to dbt users. People who spent their days in dbt should not require a second professional identity to monitor it. Salom later summarized the principle neatly: “We kept strictly to the skillset dbt users had.”
tests · lineage · freshness
signal above noise
context · owner · fix
Build where the engineers already live
Elementary released an open-source data lineage tool in September 2021. A first outside contributor appeared in November. The company joined Y Combinator’s Winter 2022 batch, launched its dbt package in March, and reached 500 GitHub stars around its April seed round. By the end of that year, the project had 30 open-source contributors and a community large enough to generate its own momentum.
The early sequence says something about Salom’s operating style. Trust was not presented as a glossy promise. Users could inspect the code, deploy it in their own environment, and fit it to an existing project. Logs, metadata, and metrics stayed close to the warehouse. The technology tried to earn adoption by reducing the distance between a problem and the person able to understand it.
Salom also became a critic of testing for its own sake. Teams commonly had too few tests or far too many. The second condition was more treacherous because it resembled diligence. A stream of failures could turn a Slack channel into what one user called an “alerts graveyard.” Everyone learned to ignore it. A safety system that trains its users not to look is an unusually elaborate decoration.
“If you have a data stack that consumers don’t trust, is it worth having?”Maayan Salom
Her prescription was practical: track failures over time, identify flaky or redundant tests, adjust thresholds, and ask who would actually act if a model broke. Coverage mattered. So did restraint. Reliability was not the number of alarms installed in the house; it was whether the right alarm could still wake anyone.
In April 2024, Elementary introduced Elementary Cloud, an enterprise product built on its open-source foundation. The company expanded from its dbt-native opening into a broader system covering observability, quality, governance, and discovery. By late 2025, Elementary 2.0 connected SQL, Python, data science workloads, and AI pipelines through a shared context layer. Agents could investigate failures, suggest missing tests, analyze costs, and draft proposed code changes. Engineers still set the rules and approved the work.
Machine speed, human sign-off
Salom’s current argument about AI begins with a sober observation: automation inherits the organization it enters. A human analyst who sees a suspicious number may pause, summon experience, and ask a colleague whether it looks right. An agent can accept the same number and act repeatedly, quickly, and without embarrassment. Old data problems acquire new velocity.
This makes Salom an optimist with conditions. She wants agents to take repetitive operational work away from data teams: incident triage, test creation, documentation, performance review, perhaps even a proposed repair. But her preferred system attaches reasoning, waits for approval, and preserves ownership. In a July 2026 demonstration with Avidov, agents worked on schedules and triggers while the data team supervised. The future on display was not an empty office. It was a better division of labor.
The principle was tested under less convenient circumstances in April 2026. An attacker exploited a script-injection weakness in a GitHub Actions workflow and published a malicious version of Elementary’s open-source Python CLI. The team removed the release, rotated credentials, shipped a clean version, hardened the release flow, and published a detailed incident timeline with specific guidance for affected users. It also worked with Wiz on investigation and defense.
For a company selling reliability, the episode could not be dismissed as someone else’s subject. The response made trust operational: state what happened, bound the impact, show the chronology, name the repairs. Transparency becomes meaningful precisely when discretion would be more comfortable.
The useful discomfort of doubt
Salom’s public writing contains an appealing mixture of seriousness and mischief. She has compared Snowflake Summit to Burning Man for data people, with conference swag standing in for desert dust. She writes about hidden Football Manager attributes and fake news inside data stacks. The humor works because it does not soften the argument. It makes an abstract infrastructure problem recognizable.
Her career has moved through roles where a clean surface can conceal an ugly mechanism: intelligence analysis, cyber research, incident response, and enterprise data. In each, the person who notices that something is off performs a valuable social function. Doubt can look inconvenient in the meeting where everyone wants the chart to be finished. Later, it often looks like competence.
Elementary’s evolution also preserves a thread from those eighth-grade spreadsheets. Visible metrics are never the whole system. Context matters. Definitions matter. Lineage matters. The person, or agent, reading a number needs to know where it came from, whether it is fresh, what changed, and who can answer for it.
Salom’s aspiration is not a world with more dashboards blinking green. It is a world where reliable data becomes part of building, where problems surface before users discover them, and where machines carry operational weight without quietly inheriting institutional confusion. The sharp question remains the same. The machinery around answering it has grown considerably.
A number does not become trustworthy because it arrived quickly. The useful system explains where it came from, what changed, and who remains responsible.
There is a founder’s lesson in that continuity. Salom did not invent her central question after choosing a market. She followed a question that had kept interrupting her, from a football simulation to security investigations to the data warehouse. Markets enjoy new terminology. Careers are often built on older irritations.