How Conveo gave every engineer the power to investigate their own issues with Grafana Assistant

When Conveo’s engineering team was small, patchy observability was manageable. Logs lived in Cloud Run, some metrics in GCP, alerting in a separate tool — and debugging meant knowing where to look before you could start looking. As the team grew, that fragmentation became friction.

Conveo, an AI-led market research platform helping brands conduct studies and surface insights at scale, adopted Grafana Cloud to bring metrics, logs, and traces together using OpenTelemetry. As the engineering team scaled from three or four engineers to nearly 12 — with plans to reach 20 — Grafana Cloud gave them the observability foundation to keep pace. More recently, Grafana Assistant changed how engineers interact with that telemetry: today, engineers who have never written a query language can investigate incidents and build dashboards on their own.

“For me, good observability means: ‘Can you ask any question about the system and get a clear, actionable answer?’” said Ludovic Vannoorenberghe, Senior Platform Engineer at Conveo.

Ludovic recently spoke with Grafana Labs about building observability at an AI-led company, how Grafana Cloud supported Conveo through an infrastructure migration, and how Grafana Assistant is changing the way engineers interact with observability data.

Can you introduce yourself and tell us a little about Conveo?

I’m Ludovic Vannoorenberghe, Senior Platform Engineer at Conveo. I joined almost a year ago. I focus primarily on infrastructure, although I also do some product work. We’re still a relatively small engineering organization, so we’re one big team rather than having completely separate infrastructure and product teams.

When I joined, Conveo had fewer than 15 employees and only three or four engineers. Today, we’re almost 80 people with around 12 engineers, and we’re looking to grow the engineering team to 20 by the end of the year.

We’re an AI-led market research platform. AI is at the center of the product and supports the research process from study design and participant recruitment through interviews and analysis. This allows our customers to conduct many more interviews at a much lower cost than they could with traditional focus groups or market research.

What makes engineering at an AI-native company different?

The pace of change is one of the biggest challenges.

Before AI, engineering had a set of established practices that had developed over maybe 15 years. There was more of a playbook you could follow. That’s less true today.

We’re constantly trying things, seeing what works, staying curious, and keeping up with the pace of change. A lot of what we’re doing is new, so the best practices don’t necessarily exist yet.

What did observability look like before Grafana Cloud?

When I joined, we were running on Google Cloud Platform (GCP) with Cloud Run. Our logs were in Cloud Run, we had some alerting through Better Stack, and we had some basic metrics in GCP.

It was difficult to query everything. Finding logs for a specific time wasn’t straightforward, and it wasn’t clear how to query them efficiently. It was similar with metrics: We didn’t have many high-quality signals.

Better Stack was great for getting started with alerting, but we quickly reached a ceiling in terms of the more complex alerting we wanted to do.

As we hired more developers, it became clear that we needed a better way to see all the signals we were sending and have a single pane of glass. There wasn’t one specific incident that triggered the change. It was more an accumulation of frustration around debugging and diagnosing issues because it was difficult to find what we needed for a particular deployment.

Why did Grafana Cloud stand out?

I’d worked with Grafana at previous companies, both self-hosted and with Grafana Cloud, so I was already familiar with the platform.

For Conveo, having a single pane of glass for metrics, traces, and logs was important. Grafana’s commitment to open source also appealed to us. A lot of our engineering team likes open source, and we were already using OpenTelemetry, so getting buy-in was easy.

We started simply by sending metrics, traces, and logs through OpenTelemetry to Grafana Cloud.

How has Grafana Cloud helped Conveo scale its infrastructure?

We moved from Cloud Run to AWS and Kubernetes because we wanted the flexibility to operate across multiple regions and potentially run within customer environments, which was becoming more important as we worked with enterprise customers.

Grafana Cloud’s value really came through after the infrastructure migration. It helped us make sure everything was running correctly and gave us confidence in the new environment.

Setting up Grafana with Kubernetes was very easy. Once it was installed, we had the telemetry and signals we needed. There was some work to make sure everything was correlated correctly, but having the alerts and visibility in place gave us confidence that things were running as expected.

That foundation is important as we continue to scale. We know we can keep adding engineers and expanding the infrastructure without having to rethink the observability layer every time.

How has Grafana Assistant changed the way your engineers use observability?

Things have changed quite a lot.

At previous companies, I spent a lot of time training teams to use Grafana and teaching people how to query their data. There’s a learning curve to understanding query languages and knowing how to find the right signals.

Now, we interact with Grafana much more through Grafana Assistant. We don’t necessarily have to write the queries ourselves. We can ask Grafana Assistant a question, have it create the query, and send us a link to the result. We can also ask it to create a dashboard.

That means I spend less time training the team on how to query the data and more time making sure our signals are high quality and correlated correctly. Engineers can interact with the data directly without having to learn the query language first.

It has also changed how we think about dashboards. Previously, a few people had the knowledge to create dashboards, so we had a small number of highly curated ones. If you had an issue, you went to those dashboards.

Now, even if an engineer is diagnosing a single issue, they can ask Grafana Assistant to create a dashboard specifically for that investigation. It can be a one-time-use dashboard that probably wouldn’t have existed before.

How is Grafana Assistant helping your team respond to incidents?

One of the most valuable capabilities for us has been automatic investigations.

When we receive an alert, Assistant Investigations can investigate what happened and help us identify the relevant signals. There have already been several times when the investigation has been spot on, saving us a lot of investigation time.

We’ve also seen the experience continue to evolve with new capabilities. The newer workflow that generates hypotheses and works through whether they can be disproven is particularly useful because it helps cut through the noise and get us closer to the signals that actually matter.

For a small engineering team like ours, being able to move from an alert to an investigation quickly is very valuable.

What does good observability look like for Conveo today?

For me, good observability means being able to ask a question about the system and get an answer without having needed to know that question when we built the system.

If we can ask any question about what’s happening internally and get a fast, readable answer that helps us diagnose the issue, that’s good observability.

Ultimately, it comes down to this: Can you ask any question about the system and get a clear, actionable answer?

What’s next for Conveo’s observability journey?

As we add more engineers, I think we’ll continue doing more of what we’re already doing and introduce more people to these workflows.

We’re also starting to explore Grafana Cloud’s AI Agent Observability for our own AI agents. We’re excited about where that can go, particularly around evaluations on real traffic.

Observability for AI agents introduces a different challenge. With a traditional application, failures tend to be deterministic. With an AI agent, the response is nondeterministic by definition, so we need different ways to evaluate those responses and make sure our agents are behaving the way we expect.

Evaluating agents is an area where we want to invest more in. We want to make sure our AI agents are behaving correctly in the same way that we make sure every other application component is behaving correctly.

We’ve also really appreciated Grafana’s willingness to build a relationship with us. The Grafana team recently came on-site to spend time with our engineers and talk through new features and tools. We appreciate that relationship and look forward to continuing to work together.