Can you really ask a database a question in plain language and get a secure, permission-appropriate answer instantly? That's exactly what one of our recent projects set out to prove.
We sat down with Gaurav Khullar to discuss a Text-to-SQL project that lets anyone ask databases questions in natural language, just like talking to your company's best expert.
Talking to databases in natural language
Who are you and what do you do?
I'm Gaurav Khullar, a partner at Data Design. My career at Data Design started in AI advisory, later expanded into machine learning and generative AI implementations, and now I'm also focused on growing our India business.
You have extensive experience in using machine learning and AI. Is there a particular project that has stuck with you?
I've worked on many machine learning projects and now with generative AI. One that really stuck with me is a Text-to-SQL solution we recently delivered for a client.
Why that one?
It stands out for two reasons. First, the impact: for the first time, our client's end users could simply have conversations with structured data, even when the databases held billions of rows and complex table structures.
Second, the technology, because we used the latest agent-based AI methods. This made the work both challenging and rewarding.
What was the project about?
We solved an accessibility problem. Business users wanted answers from data without SQL knowledge or waiting for reports.
What kind of solution did you develop?
We built a chatbot-style interface where users can ask questions in English or Finnish. The system translates them into SQL queries, runs them against various databases, and returns the results both visually and in natural language.
The chatbot remembers the conversation history, so users can ask follow-up questions naturally. We also added role-based access control to ensure that people only see the data they were authorized to access.
From a security standpoint, everything ran in the client's own cloud environment, and the data never leaves their control. For flexibility, you can also choose which language model to use. Open-source or custom fine-tuned internal models are possible too.
What was the reception like for the solution?
The reception was positive. Starting agent-based AI projects is actually very easy. You can build a proof of concept quickly and reach about 70-80% accuracy fairly soon.
“The real challenge comes when you move from 70-80% toward 100%. That "long tail" requires a lot of work and special expertise to solve all the edge cases and make the system reliable.
In this project, we gained firsthand experience of what it takes to jump from a promising proof of concept to a production-ready enterprise solution. The client appreciated our transparency throughout the process and the extra effort we put in to reach the high 90% accuracy that enterprise customers expect.
What kinds of challenges arose in the project?
Managing expectations was the biggest one. Early success can create unrealistic hopes. We had to be clear that polishing even the final edge cases takes time and expertise.
Another challenge was tooling. Many of the frameworks and platforms you'd want for this kind of work didn't exist yet. We had to build internal tools and adapt our processes to manage an environment where even a small prompt change could have a major impact on the results.
Any other interesting observations?
On the process side, we also had to adapt. This kind of project is highly agile and extremely innovative, but also uncertain. Managing the environment required us to change our working methods and design processes that could handle the unpredictability.
“It was an important lesson: innovation projects like this require not just new technology, but new ways of working.
Could the same solution concept be applied elsewhere?
Yes. We've actually productized the solution. We now offer a more general self-service Text-to-SQL product where clients can set up their own environment and connect their own database to the solution. It already supports multiple relational databases and can configure metadata descriptions, business rules, safeguards, and role-based access control.
Once the setup is done, end users can move straight into the chat-based interface. We designed it to be as seamless as possible, running entirely in the client's own cloud or environment, without connecting to external systems if they prefer.
What was the best part?
For me, the best part was the learning. Not just the technical lessons, but also the process insights: an understanding of what kind of governance practices are needed to control these noisy, unpredictable environments where even a small change can have a big impact.
Those lessons have been invaluable. They made it possible to create an innovative solution, and they now help both us and our clients build better and more durable solutions going forward.
