Kaelio wants your AI to stop being confidently wrong

A data agent writes a SQL query. It compiles. It runs. Not a single error comes back. The dashboard fills in with a clean, confident number - and the number is wrong. Anyone who has run a data team knows this exact feeling, the small drop in the stomach when a report looks fine and isn't. Kaelio, a San Francisco company that went through Y Combinator in 2025, was built around taking that feeling seriously.

The founders' argument is narrow and a little contrarian: the problem with AI data agents is not that they lack access to the database. Most of them have plenty of access. The problem is that they lack context - the accumulated, usually undocumented knowledge about what a metric actually means, which tables to join, what "active user" excludes this quarter, and where the joins quietly double-count. Give a language model a schema and a question and it will happily generate SQL that looks correct. Looking correct is the trap.

"Context is infrastructure. Reliable agents need more than database access. They need durable context that every team can inspect, improve, and govern." Kaelio, company statement

What Kaelio actually builds

The company's core product is ktx, an open-source context layer released under the Apache 2.0 license. It runs locally through a command line and over MCP, the Model Context Protocol, so an agent can query it as a tool. ktx splits a company's knowledge into two halves that serve two different readers. Markdown wiki pages hold the human-readable business context - definitions, conventions, jargon, the gotchas. YAML files hold the executable half: tables, grain, joins, measures, dimensions, filters, and filter groups.

The mechanical shift is the interesting part. Instead of an agent inventing SQL from a raw schema, it asks ktx for a specific metric with the dimensions and filters it wants. ktx's planner works out the join paths, watches for problems like fanout - the silent row-multiplication that inflates a sum - and compiles the warehouse SQL itself. The model stops being the thing that guesses at correctness. The context layer becomes the thing that enforces it.

The problem, stated plainly

Most of what makes a query correct is unwritten. It lives in the head of one or two senior analysts. When they leave, or just get busy, the definitions drift, the caveats get lost, and six months later nobody fully trusts the dashboard. That fragility was tolerable when a human wrote every query and could sanity-check the output. Point an autonomous agent at the same ambiguity and it will amplify it - fast, and with total confidence.

Kaelio's response is to make the context durable and executable at the same time. The Markdown explains the rule to a person; the YAML runs it for a machine; and both are version-controlled files a reviewer can read, correct, and approve. Documentation that actually does something, rather than a stale wiki page nobody opens.

There is a second-order benefit that data leaders tend to notice quickly. Once the definitions live in files, they can be reviewed like code. A change to how revenue is counted becomes a diff someone approves, not a quiet edit buried in a query. When a finance lead and an analytics lead disagree about what "churn" means, the argument happens once, in the open, and the resolution gets written down where the agent will actually use it. The context stops being folklore and starts being a shared contract.

It also changes what "trust" means in practice. An answer from a Kaelio-backed agent can carry its receipts: the approved definition it used, the source it read, the path it took to get there. That traceability is the difference between a number you paste into a board deck and a number you have to go re-derive by hand because you cannot explain where it came from.

Who is building it

Luca Martial, the co-founder and CEO, came up as a data scientist and NLP engineer working on enterprise data systems and agent reliability - the exact seam where this problem lives. Before Kaelio he worked on a modern pricing engine for retail. Andrey Avtomonov, the co-founder and CTO, is a repeat founder whose earlier stops include CERN and Dataiku, with a background in production AI systems and AI safety. Two people who have spent real time watching machines make plausible mistakes, now building the layer that catches them.

The AI-safety thread running through both founders is easy to miss and probably load-bearing. Safety work trains a particular habit of mind: assume the system will be wrong in ways that are hard to see, and design so that being wrong is at least visible. That is a very different instinct from the "give the model more tokens and hope" energy of a lot of the current agent boom, and it shows up in the product. Kaelio is not trying to make the model smarter. It is trying to make the model's reasoning inspectable, so a human can catch the failure before it reaches a decision.

Agents can write SQL that looks correct and runs without errors, yet still returns the wrong result. The bug Kaelio was built to fix

Products and business model

Kaelio runs an open-core playbook. The engine is free and open; the money sits in the operational layer around it.

ktx - the open-source context layer, local, over CLI and MCP. ktx Cloud - the hosted version, with multi-user governance, review workflows, SSO, continuous ingest, and observability, for teams that need control rather than a laptop tool. Data Agent - a managed AI agent grounded in that context layer, reachable in Slack, on the web, and over email, returning answers, digests, and alerts with the source evidence attached.

Plugs into the stack you already run

How it differs from the alternatives

The idea underneath ktx is not new - it is the semantic layer, the thing that powered a decade of business-intelligence dashboards by declaring, once, what a metric means. Kaelio's move is to repoint that idea at AI agents and open-source the core. That sets it apart from two neighboring camps. On one side sit closed text-to-SQL and analytics agents that hand you an answer and ask you to trust it. On the other sit semantic-layer and metrics tools - Cube, dbt's Semantic Layer and MetricFlow, LookML - built for human BI rather than for governing what an autonomous agent is allowed to say. Kaelio is trying to be the inspectable, governable context between the two.

Where it fits, and who is using it

Kaelio sits in the fast-forming "agent stack" - the infrastructure layer companies are assembling so that AI can touch real business systems without breaking them. Its natural buyers are data and engineering teams rolling out AI analytics who have already been burned once by a wrong-but-clean answer. The company has pointed to early users and logos including Mercy, Gladia, Poppins, BeSimple, Alt, Handshake, Raft, Vapi, Junction, and Maximus.

The funding is early. Kaelio raised its Y Combinator-era round in 2025, led by Rebel Fund, with Nicolas Dessaigne as its YC partner. Public trackers disagree on the exact figure - some report around $500K, others list a smaller number - which is normal for a company this young. What is clear is the shape of the bet: a small, technical team giving away the engine to earn adoption, then selling the governance and the managed agent on top.

There is a tidy lesson here even for people who will never install ktx. Write down what your numbers mean before you automate on top of them. Agents amplify whatever context you feed them, ambiguity included. Good context turns out to be the cheapest reliability a company can buy - and Kaelio is betting a company that most teams have not been buying nearly enough of it.