Tarek Gasmi

Making hidden dependencies visible

Most systems fail quietly before they fail visibly.

A country exports its groundwater without ever pricing its depletion. A technological transition appears successful until the resources it borrowed are counted. An organization believes it owns its intelligence when it is, in practice, renting it by the token.

My work is about making these blind spots visible. That often requires building the instruments first, because it is difficult to investigate what cannot be accessed, connected or queried.

Building the instruments

Many important questions go unanswered not because the data is missing, but because it is fragmented, difficult to reach programmatically and designed primarily for a reader who opens a page and downloads a file.

That reader is no longer the only audience. Public information is increasingly consulted through AI systems, yet very little of it is designed to be used by them.

Through tanitdata, I build AI-native access layers over public data: Model Context Protocol servers and query interfaces that allow statistical, agricultural and energy information to be interrogated directly by machines as well as people.

The objective is not to publish more datasets. It is to reduce the cost of asking a serious question, reproducing an analysis and contesting an established account.

The same protocol that opens a public-data portal to an agent can also connect an agent to a company's internal tools. The stakes change, but questions of permission, exposure and control remain. That is why the security of these interfaces is part of my research rather than adjacent to it: I examine the technical layer on which I also build.

I work as a research engineer rather than as an analyst who commissions tools. The instrument, the analysis and the published claim are treated as parts of the same artefact.

This means building reusable infrastructure before knowing every question it will eventually be asked to answer. That is a strategic bet, and it does not always pay. It is also the model I am testing for making rigorous public-interest research more affordable where sustained institutional funding is absent.

What the evidence reveals

Much of my empirical work is grounded in Tunisia, a setting where fragmented information, constrained analytical capacity and resource-intensive transitions make hidden dependencies unusually visible. The questions themselves extend well beyond a single country.

My research has combined more than two decades of satellite, climate, agricultural, economic and administrative data to examine groundwater depletion, agricultural production, energy exposure and technology adoption.

That work includes research on fossil groundwater in southern Tunisia's date-producing regions and on the spread of solar irrigation. Solar pumping removes an important energy-cost constraint on extraction, meaning that a transition to cleaner energy can also intensify pressure on a non-renewable water resource unless the extraction itself is controlled.

Some investigations are narrowed, revised or discontinued when the evidence does not support the original claim. That is part of the value of building analytical instruments capable of challenging the assumptions that motivated them.

Inside the systems I build

At the institutional scale, the same blindness is more immediate. An organization can hold substantial data and deploy capable AI systems while having no clear view of how those systems reach its internal tools, where permission and responsibility actually sit, or what it now depends on to keep working.

Through DataDoIt, I build enterprise generative AI systems, agentic workflows and integrations with live organizational processes. This makes me one of the people creating the dependencies I study: what a model can reach, what it is allowed to do, which external services it relies on and who can detect when something goes wrong.

Governance, for me, begins at the level of system design: who is allowed to do what, through which interface, with what visibility and under whose responsibility.

Policies become meaningful only when they correspond to the actual technical pathways through which models access data, invoke tools and take action.

Importing intelligence

At a wider scale, I am developing Importing Intelligence as a framework for examining how countries and organizations acquire externally produced AI capability through inference rather than through the domestic production of frontier models.

The inquiry began with a practical observation: access to advanced AI is not determined by technical demand alone. In the initial framework, I mapped seven points at which access can be constrained—eligibility, supported-country rules, payment rails, rate limits, request routing, contractual terms and jurisdiction.

An organization can build critical processes around an external model while having limited control over its availability, pricing, permitted uses, data routes or future contractual conditions.

The question is not whether every external dependency can be eliminated. It is which dependencies are acceptable, which are strategic and which remain invisible until access is restricted or interrupted.

Developing capacity

Teaching and academic supervision are not separate from this work. They are where questions of infrastructure and dependence become practical: who can build these systems, who can inspect them, who can challenge their outputs and who remains dependent on tools they cannot understand or modify.

Developing local capacity does not require reproducing every layer of the global technology stack. It requires the ability to make informed choices about what to build, what to import, what to expose to external systems and what must remain under direct institutional control.