From a crop-monitoring centre to an insurer's technology chair

Mandate

Replace the agricultural core system, bring remote sensing into loss assessment and strengthen data protection after the Data Security Law and the Personal Information Protection Law took effect in 2021.

Outcomes

• Replaced village-level agricultural records with a ledger of 4.6 million households and 38 million plots, cutting subsidy reconciliation from 74 days to 19.
• Paid RMB 714 million after the June 2024 harvest rain with a median of 29 days, using satellite strata checked against 1,840 witnessed sample plots.
• Cut the small-to-large farm flag-rate ratio of the sow-claims fraud model from 3.4 to 1.3, while confirmed recoveries rose to RMB 51 million in 2024.
• Moved drone imagery processing into a wholly owned subsidiary after finding contractor copies of household identity data, and classified 1,140 data sets by sensitivity.
• Reduced recovery time for the agricultural ledger from 19 hours in the 2023 test to just over three hours in 2024.

Decisions in context

Hesui's president asked me to join in March 2021 to rebuild the systems behind its agricultural business, then 38 per cent of premium and run on a platform designed in 2009 for village-level group policies. Four months later the July 2021 floods in Henan tested that platform to its limits, and agricultural claims that year took a median of 63 days to pay. I set three priorities: a ledger built around households and plots, remote sensing used with witnessed sampling rather than instead of it, and claims models that route rather than decide. I lead 610 technology staff in Zhengzhou and Hangzhou with about 200 contractors and nine direct reports, and I am a director of the wholly owned subsidiary that runs our drone survey teams. I present to the board's Risk Management and Consumer Rights Protection Committee every quarter and report to the Audit Committee on technology controls. The motor claims platform I deferred in 2022 completes in 2027.

Reason for the move

Current role. Not seeking an executive move; considering one independent directorship with the consent of Hesui's board.

Mandate

Run the issuance, claims-automation and data platforms through two years of rapid growth and set the rules for data received from platform partners.

Outcomes

• Sustained a peak of 41,000 policies issued per second on the busiest day of the November 2020 shopping festival without a failed sale.
• Raised straight-through settlement of return-shipping claims from 64 to 92 per cent using carrier tracking data.
• Declined buyer-level return histories from a platform partner and priced on merchant and category data, accepting a 3.8-point rise in the loss ratio.
• Recovered RMB 23 million in 2020 through merchant-collusion detection.
• Cut cloud cost per thousand policies by 37 per cent over two years.

Decisions in context

As deputy CTO I was responsible for the three platforms that made Yunzhu work: policy issuance, claims automation and the data platform that handled feeds from more than forty platform partners. Volume was the obvious challenge. On the busiest day of the November 2020 shopping festival we issued 41,000 policies a second without a failed sale, and over two years we brought cloud cost per thousand policies down by 37 per cent. The less obvious challenge was deciding what data to accept. When a large partner offered individual buyers' return histories for pricing in 2019, I argued for declining them and pricing on merchant and category instead. The loss ratio rose 3.8 points in 2020, and I spent much of that year explaining why to the board's risk committee. I left when Hesui offered me the chance to return to agricultural risk, the field where my career had started.

Reason for the move

Invited to lead technology at an agricultural insurer, returning to the farm risks on which I had spent my first decade.

Mandate

Rebuild policy issuance, which still ran on overnight batch files for some partners, as a real-time service.

Outcomes

• Moved 23 platform integrations from nightly batch files to real-time issuance.
• Gave every micro-policy a stored wording version and a record retrievable without the partner, closing a gap where some partners held the only copy.
• Reduced issuance time at the 99th percentile from 1.8 seconds to 140 milliseconds.

Decisions in context

I joined Yunzhu in Hangzhou in June 2017 to rebuild its policy-issuance platform. The insurer was three years old and growing fast, and it still issued policies for some platform partners through files exchanged overnight, so a customer could buy cover that we did not record until the next morning. With a team of 140, I moved 23 integrations to real-time issuance, cut issuance time at the 99th percentile from 1.8 seconds to 140 milliseconds and gave every policy a stored wording version and a record the customer could retrieve without asking the platform. Those records mattered more than the speed. They meant that a one-yuan policy was treated as an insurance contract rather than as a line in a partner's order system, and they made the later automation of claims possible. I was promoted to deputy CTO in April 2019.

Reason for the move

Promoted to deputy chief technology officer.

Mandate

Bring photo-based handling to small motor claims and replace paper survey forms in agricultural claims.

Outcomes

• Launched photo-based handling for motor claims under RMB 10,000, used for 46 per cent of eligible claims by 2016 and cutting their average cycle from nine days to two.
• Built the first rules-based anti-fraud engine for motor claims, recovering RMB 130 million in its first full year.
• Equipped 3,400 agricultural surveyors with a mobile app recording location, photographs and witness signatures.

Decisions in context

Jinhuai was my first role inside an insurer. I led claims technology at its Beijing head office, with 95 engineers and a functional line to the head of claims. Motor claims were the volume problem: hundreds of thousands of small claims a year, each of which had needed a surveyor's visit. We built a process in which policyholders photographed the damage on a phone and a desk assessor settled claims below RMB 10,000 from the photographs. By 2016 it handled 46 per cent of eligible claims and had cut their average cycle from nine days to two. We also built the company's first rules-based motor anti-fraud engine. For agricultural claims, I replaced paper survey forms with an app recording location, photographs and witness signatures, used by 3,400 surveyors. I left in 2017 to learn how to build insurance systems for the volumes that only internet sales produce.

Reason for the move

Wanted to learn to build insurance systems for the volumes that only internet sales produce.

Mandate

Turn the company's land-mapping products into plot-level tools for crop insurers.

Outcomes

• Delivered plot-boundary mapping for crop insurers in 62 counties across 14 provinces.
• Developed a method for matching mapped plots to households' land-contract records, reducing unmatched plots in pilot counties from about 30 per cent to 7 per cent.
• Grew insurance-solutions revenue from nothing to RMB 46 million a year.

Decisions in context

Lanjing sold geographic information software to land administration and agricultural bureaus, and wanted to sell to the crop insurers that had begun buying planted-area maps. I joined as a product manager and became head of insurance solutions in 2011, leading a team that grew to 40. Over five years we mapped plot boundaries for insurers in 62 counties across 14 provinces. The technical work was manageable. The difficult part was linking each mapped plot to the household that held its land-contract rights, in villages where records were on paper and boundaries had moved. Working with village committees, we developed a matching method that reduced unmatched plots in pilot counties from about 30 per cent to 7 per cent. Insurance revenue grew from nothing to RMB 46 million in 2012. I left to work inside an insurer, where the claims decisions our maps informed were actually made.

Reason for the move

Wanted to work inside an insurer, where claims decisions were actually made.

Mandate

Produce satellite-based estimates of winter-wheat area, condition and yield for provincial agricultural planning.

Outcomes

• Built a winter-wheat condition index from satellite imagery for 108 counties.
• Reduced the error of county yield estimates against final statistics from about 9 per cent to 5 per cent.
• Prepared the first planted-area maps the centre supplied to crop insurers, in 2007.

Decisions in context

After my master's degree I joined the Xilang centre in Zhengzhou, a provincial public institution that monitored crops for agricultural planning. I built satellite-based condition indices for winter wheat across 108 counties and, from 2006, led a group of six producing yield estimates each spring. We reduced the error of county yield estimates against final statistics from about 9 per cent to 5 per cent, mostly by learning where irrigation and soil made a county's average misleading. In 2007, as central premium subsidies for crop insurance began in a first group of provinces, insurers started asking the centre for planted-area maps, and I prepared the first ones we supplied. Those requests showed me that insurers needed plot-level data that our science did not yet provide. I left in 2008 to build it at a software company.

Reason for the move

Joined a software company to build the plot-level tools that crop insurers had begun asking for.