How Neho rebuilt its data platform on Google Cloud for self-service and AI

Neho is a real estate agency based in Nyon, Switzerland, and one of the country's leading agencies by number of sales. Co-founder and CEO Eric Corradin started the company in 2018, and it now employs 140 people.

Neho changed the way real estate is sold in Switzerland by charging sellers a flat fee instead of the traditional commission. It is also a hybrid player: it relies heavily on technology, with its own client platform and CRM for realtors, while its agents work on the ground with sellers and buyers every day.

Data runs through the whole business. Finance, HR, the B2B and B2C sales teams, the mortgage brokering department, marketing and product all depend on it, and it even feeds the company's bonus and incentive schemes.

Like most young companies, Neho started small with data. The data department began with a single person, and the company avoided heavy spending on tools, servers and infrastructure. Over the years, several analysts joined and replaced one another, and the stack grew one layer at a time.

Eventually, Guillaume Cougard, Head of Product & Data, saw that the foundations were not sound and that data was very hard to scale across the organization. The two people on the data team spent their days on daily requests, bug fixes and explaining what the data meant. "They were just reactive to small requests and small problems," he says.

Mahmoud Sellami, Data Engineer and Data Scientist at Neho, puts it more bluntly: "We were a fire department and not a data department. We spent most of our time fixing errors here and there instead of building new insights, new value for our company."

Fragile pipelines and a dbt monolith

The problems started at ingestion. Imports ran on fragile, long Python scripts that broke for no apparent reason, and a full run took at least an hour. Every time a developer changed something as small as a column name, the whole pipeline had to run again: another hour of waiting.

The transformation layer was a single dbt project with more than a thousand models. Some models had more than a thousand lines of code, with no documentation and heavy interdependencies.

When someone in marketing renamed a column in a Google Sheet that a dbt model read from, the model broke, and the whole company was left without fresh data.

Dashboards no one could read

Without documentation, dozens of Looker Studio dashboards had no owner and no metric definitions. Understanding a chart meant reading its code, so business users had to ask the data team. "For some task that takes 30 seconds, they would need to create a ticket, and if they were lucky we would answer in one week," Mahmoud recalls.

When data feeds incentives, trust is everything

For Eric, Neho's co-founder and CEO, the trigger came from every part of the company: a growing mistrust and frustration with data. Because Neho uses data for bonus and incentive schemes, wrong numbers hit operational teams directly.

Neho wanted more consistent, more reliable data, and more autonomy for the people using it. Hiring was not the obvious answer. Eric explains that the company had neither the expertise nor the seniority in-house, and nobody on the team had gone through a transition of this kind before.

Building a complete data team was not the right move at that stage either. Neho prefers a step-by-step, pragmatic approach to problems, so external help was the natural route. “We didn't have the skill internally to do this properly at the scale that Astrafy did, so we definitely needed a partner to do it for us”, says Guillaume.

Neho did not settle on the first option. "We had interviews with probably three different companies, and chose you guys because it seemed you were the best fit for us," Eric says.

Neho decided to rebuild everything from the ground up. The first step was a migration from the old infrastructure, running on an on-premise server, to a more modern and stable one on Google Cloud.

After an initial phase of discussions and analysis of Neho's situation, Astrafy spent significant time setting up the bedrock for the data foundations. Neho's team did not need to be hands-on in this phase. "You guys did that," Guillaume says.

Once the infrastructure was in place, the real work started on both sides. Astrafy and Neho's data team rebuilt the models and dashboards used across the company, and cleaned up many of them along the way. Getting rid of everything that had accumulated over the years and was no longer used by the business was one of the goals of the migration.

Rebuilding while the old stack kept running

During the migration, Neho's stakeholders still relied on the old stack. The data team kept it alive, fixing it only when it broke, while the new platform was built. The faster users could move to the new stack, the easier things would get for everyone.

To get there, Astrafy engineers worked on the project every day, hand in hand with Neho's data people. Astrafy brought extra workforce on top of the knowledge and technical skills needed to rebuild the entire stack.

The work was never fully delegated. Business knowledge and close contact with stakeholders were key to building dashboards the business could actually use. Guillaume describes it as a partnership: Astrafy focused on implementation and the technical side, while Neho's data team focused on user needs and refining the existing dashboards.

With the foundations in place, the team tackled the pain points Mahmoud had described one by one. "We solved most of our pain points," he says.

Ingestion: from 1 hour to 4 minutes

The biggest change was ingestion. The fragile Python scripts were replaced by Airbyte with change data capture (CDC) activated. A full import went from 1 hour to just 4 minutes.

Because Airbyte relies on public, well-maintained connectors, Neho no longer maintains any import code itself. The sudden failures disappeared, and ingestion now runs smoothly.

One data product per department

The dbt monolith was reworked into separate data products, roughly one per department, each with its own dbt project. Departments are now isolated: if something breaks in one, the others keep running. "It made our infrastructure more robust," Mahmoud says.

Documentation as a requirement

Astrafy insisted that every model have its own YAML file with its definition and documentation. That documentation later made it possible to build an MCP server on top of Neho's data.

Neho had no semantic layer before working with Astrafy. Today, metrics have shared definitions across departments. Previously, one department could use one definition for a number and another department a different one.

Now everyone knows what a metric means and who is responsible for it. Any change to a definition requires the departments to find common ground first. For Guillaume, that clarity is one of the biggest changes the migration brought. The business teams' needs stayed the same; what changed is how much they can rely on the answers.

Guillaume sees the semantic layer as essential for more than dashboards. "It's not only for the end users," he says. "I think it's mostly, today, for AI models."

Conversational analytics through an MCP server

With Astrafy's help, Neho also built an MCP server so employees can use an LLM to ask questions about their data in natural language and get immediate answers. When Neho started the project with Astrafy, it wasn't yet clear whether conversational analytics would be possible. Once the API was available, the MCP server was built in a week.

The MCP server is still in a pilot phase, but feedback has been very positive. Users are more autonomous and can get insights even when the data is not on a dashboard. Mahmoud is clear about why it works.

Guillaume expects both formats to coexist. Dashboards remain useful for recurring daily checkups of known metrics. Custom, one-off questions will go to conversational analytics. "You cannot do that if your data stack is not sound, if you're not relying on reliable data," he says.

Astrafy worked with Neho on a time-and-materials basis: Astrafy engineers were embedded in the team and worked alongside Neho's data people every day, rather than delivering a fixed scope over a set number of months.

Continuous learning, continuous delivery

The team worked in sprints with fixed objectives, but talked every day. "I would call it CL/CD: continuous learning, continuous delivery," Mahmoud says. Astrafy engineers were "just one Slack call away" whenever the team needed help.

The working pattern was consistent. Neho presented a pain point, Astrafy proposed a solution and walked through current best practices, and then both teams coded the first piece together.

After that, Neho's team carried on independently. Mahmoud says he learned a lot, from best practices and new tools to governance, both professionally and personally.

Fit for a company of Neho's size

Astrafy did not apply a one-size-fits-all solution. "They did not just deliver something generic," Mahmoud says. "When we thought a tool was overkill for a company of our size, they would simplify it or even drop it."

Autonomy as the goal

Astrafy brought new technologies that Neho's team wasn't used to, so the team had to adapt and learn. According to Guillaume, the team could rely on Astrafy for every complex case or unfamiliar feature, because Astrafy knows these tools well.

“My team learned a lot from them and is now very capable of handling our data infrastructure autonomously."

Guillaume Cougard, Head of Product & Data, Neho

“My team learned a lot from them and is now very capable of handling our data infrastructure autonomously."

Guillaume Cougard, Head of Product & Data, Neho

For Mahmoud, this focus on autonomy is what sets Astrafy apart.

“They were making the effort to make us as autonomous as possible, which is not what every consultancy company does. Most of them would make themselves indispensable, which I don't think was the case for Astrafy."

Mahmoud Sellami, Data Engineer & Data Scientist, Neho

“They were making the effort to make us as autonomous as possible, which is not what every consultancy company does. Most of them would make themselves indispensable, which I don't think was the case for Astrafy."

Mahmoud Sellami, Data Engineer & Data Scientist, Neho

Guillaume describes the collaboration in similar terms. "We felt in good hands with capable, technically very skilled people that really wanted to understand how we work," he says. "It wasn't like relying on an external stakeholder and just waiting for answers. It was very fluid and very nice."

Neho measures the success of its data team in a simple way: it asks stakeholders whether they are more independent and whether they can rely on the data. On both counts, Guillaume has seen a lot of improvement.

Stakeholders now answer their own questions and build their own dashboards. The data team is no longer answering ticket after ticket and has time for projects that add real value. One example is a model that predicts when, and by how much, Neho should reduce the price of a property that is not selling well. In the past, the team had no time for this kind of work.

For Eric, the first real outcome is that the numbers are right down to the details. That matters most for leaders who manage people and incentive schemes. He considers the objective accomplished, or well on its way.

“Having really solid data foundations means that we can go to the next steps and create really interesting strategic projects. Without that, it would not be possible."

Eric Corradin, Co-founder & CEO, Neho

“Having really solid data foundations means that we can go to the next steps and create really interesting strategic projects. Without that, it would not be possible."

Eric Corradin, Co-founder & CEO, Neho