Mingsheng Hong is an AI and data-infrastructure executive at Ironclad. He announced that he was joining Ironclad as Head of AI, and in his 2026 AI Engineer World’s Fair talk described himself as a vice president of engineering focused on AI. His work at Ironclad centers on AI-native products and features for legal contracting.
Event-stream research and database products
Hong received his Ph.D. in computer science from Cornell University in November 2008. Working with Johannes Gehrke, he studied expressive, large-scale event-stream processing and helped build Cayuga, a system for detecting patterns over event streams. During several summers at Microsoft Research, he contributed to the CEDR event-processing and Immortal DB transaction-time database projects. His Cornell homepage lists his thesis, Expressive and Scalable Event Stream Processing, and publications on Cayuga and related database research.
After completing his Ph.D., Hong joined Vertica. He later recalled leading the team that built Database Designer, an automatic tuner for Vertica, and moving into customer-facing work that included product evangelism and enterprise sales. At Hadapt, he participated in customer discussions about using SQL on Hadoop to obtain faster analytical insights at lower cost. These experiences broadened his work from database implementation to explaining technical products and understanding customer requirements.
Data and machine-learning infrastructure at Google
Hong joined Google in 2013 and later said that he spent 8.75 years there working on data and machine-learning infrastructure. He led or contributed to systems including Mesa, Napa’s columnar storage and querying, TensorFlow Eager, Mesh TensorFlow, and TFRT. He described leading a migration of an advertising data backend from row-oriented to column-oriented storage while preserving service for existing workloads.
Hong served as TFRT’s technical lead and manager. In a 2020 TensorFlow introduction, he and product manager Eric Johnson described TFRT as a low-level runtime intended to execute TensorFlow workloads efficiently across varied hardware. Its design emphasized asynchronous execution, efficient use of multithreaded host CPUs, modular device runtimes, and support for eager and graph execution. A 2022 progress update reported deployments across Google training and inference workloads and discussed lessons involving asynchronous and synchronous execution, compiler optimization, and cost-based scheduling.
Bluesky Data and legal AI
After leaving Google, Hong co-founded Bluesky Data with Zheng Shao and served as its CEO. The company worked on making cloud data workloads more cost-efficient through monitoring, analysis, and optimization recommendations. Hong described its early product work with Snowflake users, including recommendations involving queries, data layout, and warehouse settings. His account of Bluesky emphasized measuring the business value produced by cloud computation rather than treating infrastructure spending as an isolated number.
At Ironclad, Hong applies a related concern with measurable value and trust to legal contracting. He has described contracts as central to hiring, enterprise sales, partnerships, and fundraising, while also creating significant friction for businesses. In his 2026 talk, “From Tokenmaxxing to Trusted Throughput,” he explained that lawyers may first test conversational search against contracts they already understand. Results that match their expectations can support later use on unfamiliar agreements or workflows such as redlining and anomaly detection.
Measuring useful AI output
Hong calls reviewed and validated AI-assisted engineering output trusted throughput. The framework combines three forms of evidence:
automated checks such as tests, security checks, and canary rollouts;
human judgment through code and design review, including assessments of clarity, maintainability, and architectural fit; and
customer evidence such as production rollbacks, support tickets, bugs, and usability complaints.
Under this framework, generated or merged code is not automatically valuable. Hong argues that teams should measure delivered value alongside token cost before trying to reduce spending. Token dashboards can expose adoption gaps or unusual usage, but leaderboards risk rewarding consumption rather than useful outcomes. Comparisons also require context because infrastructure and interface teams may use AI differently.
Hong has described Ironclad’s evolving attempt to quantify output. The company moved from counting opened pull requests to counting merged pull requests, then began weighting merged requests with AI-assigned, T-shirt-sized complexity scores. He presents this as a pragmatic proxy that remains incomplete: a short concurrency fix may require more consequential work than a large block of boilerplate, but complexity is not itself a settled measure of customer value.
Review, CI, and cost controls
Hong argues that abundant code generation shifts bottlenecks into code review and continuous integration. He favors AI review as a first pass for routine issues such as coding style and missing tests, allowing human reviewers to concentrate on architecture, maintainability, and security while retaining final accountability. He cautions that slow CI can encourage oversized pull requests, which spread reviewer attention thin and reduce review quality.
For CI, Hong recommends measuring the elapsed time from a pull request being ready to its submission or merge, along with the number of retries required to pass tests. Repeatedly rerunning flaky tests—whether by a person or an agent—consumes time or tokens without correcting the underlying instability. His team has invested in reducing flaky tests and improving CI infrastructure.
His operating controls include budgets and quotas, anomaly alerts, and finite limits on agent loops that repeatedly modify code and rerun tests. He also recommends arranging prompts with stable instructions first to improve prompt-cache reuse and summarizing or pruning accumulated context during long sessions. For build-versus-buy decisions, he favors buying common, non-differentiating infrastructure while developing internal prompts and playbooks that encode company-specific engineering practices.
Hong also treats engineers’ professional experience as part of AI adoption. Engineers who valued writing code by hand may find reviewing poor machine-generated output less satisfying. He argues that leaders should take that concern seriously and help engineers find consequential technical work through which their judgment and skills can continue to develop.
Mingsheng Hong explains how Ironclad connects AI coding costs to software that survives automated checks, human judgment, and customer use—and why faster generation moves the real bottleneck into review and CI.
Use token dashboards to investigate adoption gaps and unusual bursts, with comparisons grounded in each team’s work. Do not reward consumption.
Trusted throughput combines objective checks, human engineering judgment, and evidence from customer use. A merge is only an intermediate delivery measure.
Expect abundant generation to move constraints into review and CI. Measure readiness-to-submission time and retries, remove flaky tests, and resist giant batched PRs.
Buy common infrastructure and build company-specific working knowledge, while evaluating ambiguous agent investments in light of their downstream review and CI costs.
▶ Watch ↗