Zhao Nianxi (赵念溪) could have stayed in the lab. As an AI engineer at the Railway Fourth Survey and Design Institute (铁四院), he had access to powerful servers, clean data, and comfortable working conditions. But he believed something fundamental was missing: AI that only works in laboratories doesn't solve real problems.
He wanted to take AI to construction sites — into beam yards, tunnels, ports, prefabricated component factories, mountain roads where the network barely works.
"AI can't just perform well in the lab — it has to work on construction sites too," he said. "It's not about sitting in an air-conditioned server room. It's about going into the mountains, onto the production lines, into the field."
The first challenge was the "last mile" of deployment. Network was spotty in mountainous areas, cloud-based AI kept lagging. Camera brands were mixed — old and new — making video streams hard to unify. Zhao didn't blame "insufficient computing power." He took the problems apart piece by piece.
For unstable networks, he moved AI computing to the edge, letting analysis happen on-site. For multiple incoming video streams, he optimized scheduling and resource usage so a single device could support 50+ camera feeds simultaneously. For device connectivity, he built a local server platform supporting 100+ devices stable online.
With the equipment connected, AI still had to learn to read a construction site. General models can recognize people, cars, fire, and smoke — but they can't judge construction process status, detect water seepage, or identify on-site hazards specific to railway building.
Zhao collected 250,000 on-site images himself. He didn't wait in an office for data to come to him. He followed projects to the field, gathered requirements, captured and annotated real photos, then taught the model construction standards and risk points, step by step. Gradually, the system learned to warn about unfastened safety harnesses at height, illegal crane lifts, hot work violations — all with over 95% accuracy. Process identification and water leakage monitoring could be handled by the system as an auxiliary tool.
When fixed cameras couldn't cover blind spots, he integrated AI with drones and robot dogs, sending inspections into complex areas and dangerous locations.
Then he went further. If AI works well on one site, why rebuild it every time? He led the creation of the institute's AI general capability open platform — packaging OCR recognition, visual inspection, and large model Q&A into unified interfaces. He also participated in the integrated construction-management-operations platform development, winning silver at the 11th "Chuangqingchun" national competition.
For Zhao, the most rewarding part isn't the awards. It's watching AI become something practical — an extra set of eyes on the construction site, catching one more safety risk, preventing one more accident. He leads an AI capability group now, sharing edge deployment and large model application experience, bringing young colleagues along.
"This generation of transportation youth is lucky," he said. "We can connect code to a bridge, a tunnel, a railway — and to the pulse of our country's transportation development."