Measuring loss from a distance, and persuading the person who suffered it

I grew up in a wheat county in eastern Henan, where my father was a technician at the township agricultural station. Each June he estimated the harvest by laying a wooden frame, one metre square inside, on sample plots and counting the ears within it. Farmers often disputed his figures, because those figures mattered to them. He answered every dispute the same way: he asked the farmer to crouch beside him and count. I carried that frame through several summers. Only much later did I understand that it had taught me the central problem of my working life, which is how to measure a loss from a distance and still persuade the person who suffered it.

I studied geographic information science at Beiyuan University of Science and Technology, graduating in 2000, and took a master's degree in cartography and remote sensing at the Hanqiao Institute of Geographic Sciences in 2003. My thesis estimated winter-wheat yield from satellite imagery across twelve counties. The model looked convincing at county level and was badly wrong in individual villages where irrigation differed from the norm. That gap, between an average that is right and a household that is wrong, has followed me ever since.

From 2003 to 2008 I worked at the Xilang Agro-Meteorological and Remote Sensing Centre in Zhengzhou, a provincial public institution, producing crop-condition indices and yield estimates for 108 wheat and maize counties. In 2007, when the central government began subsidising crop-insurance premiums in a first group of provinces, insurers started asking the centre for maps of planted area. I prepared the first of them and realised that insurers needed something our science did not yet provide: not the area of wheat in a county, but the field that belonged to a particular household.

I joined Lanjing Geoinformatics, a private software company in Beijing, in 2008 to build that. Over five years my team mapped plot boundaries for crop insurers in 62 counties. The hardest part was not the imagery. It was matching a mapped plot to the household that held its land-contract rights, in villages where records were on paper, boundaries had shifted and one family farmed land leased from three others. I spent more time in village committee offices than at a screen, and learnt that good agricultural data is a social achievement before it is a technical one.

In 2013 I moved inside an insurer for the first time, as director of claims technology at Jinhuai Property Insurance in Beijing, then one of the larger property and casualty insurers in China. We built photo-based handling for small motor claims, a rules-based anti-fraud engine and a mobile survey app that gave 3,400 agricultural surveyors location stamps and witness signatures in place of paper forms. I learnt there that a claims system is judged by its worst week, not by its average one.

In 2017 I went to Hangzhou to run platform engineering at Yunzhu Online Property Insurance, an internet-only insurer selling return-shipping, delivery-delay and device cover through e-commerce and travel platforms. Premiums were often below two yuan and volumes were enormous; on the busiest shopping day of 2020 we issued 41,000 policies a second. As deputy CTO from 2019, I also had to decide what data we would accept from platform partners. Declining a feed of individual shoppers' return histories cost us loss-ratio points and taught me that minimising data is a pricing decision as much as a legal one.

Hesui's president asked me to join as CTO in March 2021. I wanted to return to farm risks, and Hesui needed its agricultural systems rebuilt. Four months after I arrived, the July 2021 floods in Henan put those systems under a load they had never been designed for, and agricultural claims that year took a median of 63 days to pay. Since then we have rebuilt the agricultural core around households and plots, brought satellite and drone imagery into loss assessment with witnessed sampling as its check, and taken our livestock fraud model out of the decision to hold a claim.

I lead 610 technology staff in Zhengzhou and Hangzhou, with about 200 contractors. I ask two things of every team. Engineers who build claims systems spend at least a week a year with surveyors in a county during harvest or after a storm, because a requirement written in Zhengzhou rarely survives a village. And every model we release has a named owner who can explain its decisions to a branch manager and, if asked, to a farmer. I report technology and data risks to the board's Risk Management and Consumer Rights Protection Committee in plain language, near misses included.

Within Chinese agricultural insurance I am probably known for three positions: that a remote-sensing loss method should be published before the season, that a fraud score should route a claim rather than refuse it, and that an insurer selling through a platform should take less data than it is offered. I am now preparing to bring those habits to a board outside insurance as an independent director.

Open to Board appointments · Speaking invitations

Languages / Mandarin Chinese · English