The AI Searchlight for Manufacturing Downtime
In manufacturing, machine downtime is a direct hit to the bottom line. Every minute a production line is stalled costs real money, and troubleshooting can be a slow, painful process. Kevin Fort and Atheesan Murgesu from IntoData recently broke down a project with lighting manufacturer Niko that tackled this problem head-on. They detailed how they helped Niko shift from messy, manual searches for information to an AI-powered system that turns hours of downtime into minutes of productivity.
The Three-Hour Problem
Here's the situation technicians at Niko used to face: a production machine breaks down, and it's their job to fix it, fast. The catch? The solution was likely buried in one of two places: a database of 88,000 old maintenance tickets filled with internal jargon and typos, or over 10,000 pages of dense technical manuals. As Kevin explained, what should have been a quick fix often became a three-hour hunt for information, bringing production to a standstill. Niko’s existing ticketing system was a chaotic library of unstructured text. The company knew it had a costly problem, but without any in-house AI specialists, it needed a partner to navigate a solution. This wasn’t a theoretical exercise; it was a daily operational headache.
Building a Searchlight, Not a Black Box
Instead of just layering a generic AI model over the existing mess, the team took a more deliberate, multi-step approach. Niko’s main requirement, as Atheesan pointed out, was for a system they could understand and control: a transparent "searchlight," not a "black box" that just spits out answers.
The first step was to build a controlled search system designed to handle Niko’s messy data. This tool allowed users to fine-tune searches, adjust the importance of results from past tickets versus technical manuals, and even switch between different search algorithms.
Next, to actually measure success, they established a "ground truth." Niko's own experts curated a set of common problems with verified, correct answers. This gave the team a clear benchmark to test against. As Atheesan noted, only with this baseline could they tell if any changes were actually making the system better. A simple "thumbs up/down" feedback button was also integrated, allowing technicians to continually refine the results.
Only once that foundation was solid did the team bring in a large language model (LLM). The LLM's job wasn't to find the answer, but to summarize the highly relevant results the search tool had already surfaced. It would then present a clear, easy-to-read solution to the technician, complete with direct links to the source documents for verification.
Empowering People, Not Replacing Them
The project’s success went beyond just the technology. Kevin stressed that the goal was never to replace technicians but to empower them with a tool that made them faster and more effective. This created a virtuous cycle for data quality: technicians quickly realized that writing clearer, more detailed information in a new ticket would lead to better, faster answers for their colleagues in the future.
A major win was the ability to capture and retain institutional knowledge. Once a problem is solved, that solution is now indexed and easily accessible for a future incident, preventing expertise from walking out the door when an employee leaves.
The project highlights a choice Niko consciously made. As one manager put it, "AI is coming to the production floor either way. The real question is, will we take control of it ourselves or simply let it happen to us?" Niko chose to take control, and within just three months, the company had a working product delivering real value.
Conclusion
The Niko project serves as a practical guide for implementing AI effectively. It all starts with a tangible business problem rooted in daily operations, not an IT initiative. From there, the focus shifted to building a robust, measurable, and controllable search system that could handle imperfect data. The LLM was the final, value-adding layer, not the starting point. The result is a system that isn't just technologically impressive but genuinely useful, turning hours of frustration into minutes of focused work.
Facing similar data challenges in your operations? Let's discuss how we can help you find the answers.
.svg){kind=small}
