A database column has the crisp authority of something that has never attended a meeting. It says revenue, perhaps, and declines to mention whether finance, sales, and the board agree on what revenue is. An AI agent can query that column in a second. It can produce immaculate SQL, a tidy chart, and an answer that is wrong in all the important ways. Luca Martial has built his latest company around this small, expensive comedy.
Martial is the co-founder and CEO of Kaelio, a San Francisco company making context infrastructure for data agents. Its open-source project, ktx, collects the material a warehouse does not naturally explain: approved metric definitions, safe joins, business vocabulary, policies, and the knowledge scattered through wikis and analytics tools. The software makes that context reviewable by people and available to agents when they act. Its argument is pleasingly unfashionable. Access to more data is not enough. The agent must know what the organization means.
That concern did not begin with databases. Martial's working life has moved through sport science, public health, healthcare natural-language processing, and AI evaluation. On paper, the itinerary looks like several careers trying to share one boarding pass. Look closer and the same problem keeps reappearing: measurement without context is a confident nuisance.
Before the model, the measurement
At Concordia University, Martial studied sport science. He worked around strength and conditioning, including youth athletic development. Sport science is full of numbers that become absurd when detached from the person producing them. A sprint time depends on conditions. A strength ratio describes a body, a task, and a method, not a Platonic ranking floating above the gym floor. The habit of asking how a number was made would prove portable.
He continued into public health at McGill University. In 2020, as a visiting scientist with the Public Health Modelling group at the University of Cambridge's MRC Epidemiology Unit, his work focused on using street-view imagery to model active-travel behavior in lower- and middle-income countries. He also contributed to research on neighborhood deprivation and the availability of inclusive leisure programs in Canadian cities. These are questions where data, environment, and human behavior refuse to separate neatly. The map may be clean. The city is not.
Then language models offered a new instrument for working with messy reality. At John Snow Labs, Martial built custom NLP systems around clinical documents and trained models for named-entity recognition, relation extraction, text classification, de-identification, and clinical entity resolution. He later described healthcare NLP as a world of shifting vocabularies, strict privacy, uneven notes, and workflows where reliability matters more than novelty. The practical lesson was that vocabularies such as SNOMED CT, ICD-10, RxNorm, and LOINC are core infrastructure. A model that cannot speak the systems around it remains an impressive guest who never learned the address.
“Long before Kaelio, I was a healthcare data scientist.”Luca Martial, reflecting on the habits he carried into the company
At John Snow Labs he also led product development for NLP Test, an open-source library for evaluating robustness, bias, fairness, accuracy, and representation. The theme tightened: the useful question was no longer merely whether a model could perform a task, but how it failed and whether anyone had looked for that failure before deployment.
Learning to distrust a smooth demo
In 2023, Martial joined Giskard in Paris as its first product manager. He worked across engineering, design, marketing, and sales, then became Product Lead for LLM safety products. His remit included guiding customers through red-team assessments of generative-AI applications. He also helped create DeepLearning.AI's short course Red Teaming LLM Applications. If healthcare NLP taught him to respect domain language, evaluation work taught him to interrogate the polished answer.
There is a charming early clue in his public work. In 2021, Martial wrote a tutorial about building an Elon Musk-inspired Discord bot. The project combined a fine-tuned conversational model with toxicity detection and a three-strike moderation system. It belonged to a more innocent moment in generative AI, when getting a model to maintain a persona was entertainment enough. Yet even the toy came with rules, monitoring, and consequences. The conversation and its guardrails arrived together.
At Giskard, Martial worked with Andrey Avtomonov, an experienced data and AI engineer who became Giskard's CTO. The pair later founded Kaelio, with Martial as CEO and Avtomonov as CTO. Their shared resume made reliability less of a marketing adjective and more of a technical preoccupation.
A company changes its aperture
Kaelio entered Y Combinator's X25 batch in 2025. Martial wrote that reaching San Francisco felt surreal after “religiously” reading Paul Graham's essays and watching Y Combinator partner material. The company launched as an analytics copilot for healthcare organizations, promising to connect clinical, operational, and financial data so leaders could ask questions, spot risks, and act without waiting weeks for a report.
The first framing was specific to healthcare, but the underlying difficulty was broader. Every organization has data distributed across warehouses, dashboards, documentation, and the memories of people who know why a metric changed in 2022. An agent sees tables and columns. It does not automatically see the argument that produced the approved definition, the join that must never be made, or the caveat everyone remembers except the new hire.
The journey from data to a defensible answer
Kaelio widened its aperture. Today the company describes ktx as an open-source context layer for data agents. It runs locally, is licensed under Apache 2.0, connects to common warehouses and analytics tools, and serves context to agents through command-line and MCP interfaces. The company also offers ktx Cloud for teams and a managed Data Agent. The names are technical; the ambition is social. Business meaning should survive the handoff from the person who knows it to the machine answering the next question.
“Every number a data & analytics AI agent outputs is a governance decision: which definition it used, and who stood behind it.”Luca Martial
That sentence captures Martial's current thesis. Governance is not paperwork applied after the agent speaks. It sits inside the answer. Which definition of active customer did the agent use? Who approved it? Can another person reproduce the result? The good answer carries its lineage like a passport.
The open-source bet
Open-sourcing ktx made the argument inspectable. By August 2026, its GitHub repository showed more than 1,600 stars, over 500 commits, and community work spanning integrations with Sigma, Google Drive, and MongoDB. Martial celebrated those additions precisely because contributors were extending the project toward the tools their own teams use. A context layer that cannot absorb local context would be a fine joke at its own expense.
The project's octopus motif suits the job. ktx reaches into databases, semantic layers, business-intelligence tools, and documents, then presents a single surface to the agent. Yet Martial and Kaelio emphasize human review. The software can organize knowledge, detect relationships, remove duplicates, and flag contradictions. People remain responsible for settling disagreements that matter. Automation does the gathering; judgment keeps the keys.
Martial's public posts dwell on metric governance, semantic layers, connector design, and the unromantic mechanics of enterprise data. This is not the language of an AI oracle. It is closer to municipal plumbing, if the pipes also had to agree on the meaning of quarterly revenue. The restraint is the point. Reliable systems are usually assembled from definitions, permissions, reviews, logs, and the occasional person willing to say that two dashboards cannot both be right.
His path offers a useful idea for builders tempted to narrate their lives as a sequence of pivots. A career can change industries while keeping faith with one stubborn question. For Martial, the settings moved from gyms to cities, clinical notes, red-team exercises, and data warehouses. The question stayed put: what must be true before we trust the output?
Kaelio's answer is context that can be read, reviewed, versioned, and used at the moment of execution. It is not a promise that machines will stop being wrong. It is a proposal for making their answers less mysterious and the people behind the definitions more visible. In a field enchanted by agents that can do almost anything, Luca Martial is building the part that tells them what the company actually meant.