Velum Labs Wants Your Data to Stop Lying to the Boardroom

There is a small, recurring humiliation inside almost every company that calls itself data-driven. Two people pull up the same number - revenue, active users, churn - and the numbers disagree. Someone changed a definition three months ago, upstream, and nobody noticed until the meeting. The room goes quiet. That quiet is the problem Velum Labs decided to build a company around.

Velum Labs is a San Francisco startup, backed by Y Combinator, that describes itself in one plain sentence: an operating system for data quality. The longer version, from its own site, is more useful. "Velum finds where your data broke, why it broke, and fixes it. Then it writes the contract so it never breaks that way again." It is a tidy summary of an unglamorous job, and the tidiness is the point.

The company is two people - co-founders Benjamin Munoz-Cerro and Alen Rubilar-Munoz - which makes the scope of the ambition slightly comic and, on inspection, slightly credible.

The core idea

01Not another alarm on the wall

Most tools in the "data quality" aisle are alarms. They notice a table is late, a row count dropped, a column went null, and they page someone. That is genuinely useful, and it is also where most of them stop. The engineer who gets paged still has to open a dozen tabs and reconstruct, by hand, which upstream change caused the number to move.

Velum's bet is that the alarm is the least interesting part. Its pitch is a loop: detect the divergence, trace it to a root cause through the actual query lineage, propose and deploy a fix through the tools a team already uses - git, dbt, CI/CD - and then generate a data contract from the real break, so that specific failure cannot happen the same way twice.

How Velum closes the loop

01

Detect

Watches query patterns, distributions & schema drift

➔

02

Trace

Root cause via live query lineage graph

➔

03

Fix

Deploys through git / dbt / CI-CD

➔

04

Contract

Auto-writes the rule to prevent recurrence

The word the founders keep using is "contract." In data teams a contract is a promise about what a table will and will not do - its shape, its freshness, its meaning. In theory everyone wants them. In practice almost nobody writes them, because writing them by hand, in YAML, before anything has gone wrong, feels like buying insurance for a house that looks fine. Velum's trick is to wait for the fire, then write the policy that matches the exact way the house burned.

There is a quiet engineering choice underneath all of this. Rather than parse code to guess how data flows, Velum reads live production query traffic and builds the dependency graph from what teams are actually running. A static parse tells you how the pipeline is supposed to work. Watching the queries tells you how it really works, including the ad-hoc join someone wrote at 11pm that half the company now depends on. That second picture is the one that breaks dashboards, and it is the one Velum optimizes for.

Who it is for

02The people who get blamed

Velum sells to enterprise data teams, and its most telling early customers are regulated fintechs. That is a deliberate choice. In a consumer app, a wrong number is embarrassing. In a company managing a billion dollars in assets, a wrong number is a compliance event. The stakes turn "the dashboard looked right" from an excuse into a liability, and that is exactly the buyer who will pay for proof rather than a feeling.

The audience inside those companies is layered, and Velum names all three: the data engineers who fix the pipes, the data leads who answer for them, and the executives who consume the dashboards and have no idea what a broken join looks like. The product has to satisfy an engineer's need for lineage and a VP's need for a number that survives scrutiny. Bridging those two audiences - "pipeline to boardroom" - is the actual product, not a slogan.

Where it sits

03A layer, not a replacement

Velum does not ask a company to move its data. It layers on top of the warehouse and the pipeline tooling a team has already standardized on, reading production query traffic to build its picture. That is a low-friction way in: no migration, no rip-and-replace, just a control plane that watches what is already flowing.

The competition

04Detection is crowded. The loop is not.

Data observability is not an empty market. Monte Carlo, Anomalo, Bigeye, Metaplane and the open-source Great Expectations have all built businesses on noticing when data breaks. Velum's differentiation is less about detecting better and more about refusing to stop there. The claim is end-to-end: from the alert to the cause to the fix to the contract, without a human stitching the four together.

Where the tools stop (illustrative)

Detect

industry standard

Root cause

partial

Auto-fix

rare

Auto-contract

Velum's wedge

The founders

05From quantum physics to broken dashboards

The two founders met in high school in Chile and never really stopped building together. Benjamin Munoz-Cerro, the CEO, studied physics and mathematics at Stanford and spent roughly five years on reinforcement-learning methods for quantum computing at Harvard and the Max Planck Institute. Alen Rubilar-Munoz, the CTO, is a mathematician and machine-learning engineer whose background runs through geometric deep learning and in-memory computing for AI.

It is a resume that points toward a frontier-AI research lab, not a company that untangles why the finance team's revenue figure disagrees with the ops team's. That gap is worth sitting with. The interesting version of Velum is that two people who could be chasing exotic problems chose an ordinary, expensive one instead - the kind that does not make headlines but quietly costs enterprises real money every quarter.

The research background is not decoration, either. Root-cause tracing across a live query graph, and generating a contract that captures the precise class of failure that occurred, are closer to graph and inference problems than to dashboard design. A team that spent years on reinforcement learning and geometric deep learning is, at minimum, unlikely to be intimidated by the hard part. The stated values line up with that: research rigor, technical excellence, and a habit of publishing rather than hoarding. For infrastructure that engineers have to trust before they wire it into production, being legible in public is a feature.

The pivot

06The company they were, briefly

Velum did not start here. When the founders announced their Y Combinator acceptance in 2025, the pitch was different and, on paper, more exotic: "zero-trust infrastructure for AI privacy using fully homomorphic encryption." The idea was to make sensitive data - medical records, financial transactions, protected IP - usable by AI without ever decrypting it. Encrypted indexing, encrypted inference, encrypted training.

Somewhere between that announcement and the current site, the company changed its mind. It kept the instinct - enterprises that cannot trust their data cannot use it - and swapped the exotic answer for the practical one. Before you can safely compute on data, you have to know the data is correct. The pivot from cryptography to correctness is not a retreat; it is a founder team following the pain their customers actually complained about. That is usually a good sign.

The business

07Small team, infrastructure ambitions

The business model is the familiar B2B infrastructure shape: sell to enterprise data teams, layer on top of the existing stack, land through design partners in industries where correctness is non-negotiable, expand as the control plane touches more of the pipeline. Velum has described being open by default - contributing to open source and publishing research - which is both a values statement and a sensible distribution strategy for infrastructure that engineers have to trust before they adopt.

The funding is early and modest by design: a seed raise recorded around $130,000 tied to the Y Combinator round, with Aaron Epstein as the primary YC partner. What Velum is really testing is not whether it can raise money but whether two researchers can turn a universally acknowledged annoyance into infrastructure a company forgets it is running - the way it forgets its database until the day it goes down. That is the bar for this category. Nobody thanks you for the fire that never started.

Will it matter in ten years? The honest answer is that the category will. Every company is accumulating more data, more definitions, and more places for those definitions to quietly drift apart. Someone will own the layer that keeps the numbers honest. Velum is a two-person wager that it can be them.