
◦
Achieved an ultra low 90th-percentile false positive rate of 0.12% on 6k baseline websites (90k webpages).
◦
Flagged 21% of 117k Common Crawl websites and 14% of 18k websites from Bing search results.
◦
Substantiated the claim that the use of LLM text grew with hard evidence.
◦
Identified clear financial incentives for websites dominated by LLM texts, mainly ads and affiliate links.
• Surveyed usage, weakness, and extensions of the Coalition for Content Provenance and Authenticity (C2PA).
Agentic Formal Verification for Rust/Verus Jun. 2026 – Oct. 2026
Midas Intelligence Inc. (math-for-AI startup), San Francisco (Remote)
• Invented the company’s software story: achieve both development velocity and reliability via agents and Verus.
• Led and single-handedly established the company’s software division. Reported directly to the president and the CEO.
• Created Midas Lex, a tool that guides agents to write Verus specifications and proofs.
• Found bugs in popular open source Rust projects like uutils coreutils. Built benchmarks of unsolved Verus tasks.
Independent research, Internet Route Verification Apr. 2023 – May 2024
Federal University of Minas Gerais, Brazil (Remote) Advisor: Italo Cunha
• Does Internet traffic follow the rules that networks publicly declare? Conducted one of the most thorough checks,
covering all public routing policies in the Internet Routing Registry (IRR). The IRR underpins much of the Internet, is
used to guide public peering, help troubleshoot routing issue, and avoid routing misconfiguration and attacks that can
cause global Internet outages.
• Single-handedly built RPSLyzer, the 1st operational interpreter to parse 99.99% of all 7 GB of routing policies in the IRR
in the Routing Policy Specification Language (RPSL), and the 1st thorough checker for whether real, observed Border
Gateway Protocol (BGP) routes follow these policies per IETF RFC 2622 and RFC 4012.
◦
Checked a massive sum of 779M BGP routes from 60 RIPE RIS and RouteViews collectors, in 3 hr with 2 GB of RAM.
◦
RPSLyzer correctly parses intricate RPSL constructs, including advanced set operations and a custom regex.
◦
Characterized the policies of 78,701 autonomous systems (ASes). Found a wide variation in RPSL usage: several ASes
declare thousands of policies, while 53% declare none at all.
• Checked and analyzed whether observed interdomain routes match routing policies with RPSLyzer; identified common
RPSL usage patterns and usage mistakes.
◦
Found that 40% of interconnections in observed BGP routes cannot be verified due to missing policies.
◦
Incorporated CAIDA AS relationship data to improve verification coverage by 19 percentage points.
◦
Provided tooling to verify routes against the RPSL, helping improve interdomain routing security.
◦
Identified and implemented checks for 6 major explanations for why routes fail to match the relevant RPSL.
Research Assistant, Mobile Federated Learning (FL) Mar. 2023 – May 2024
The FedCampus Team, EdgeIntelligence Lab, DKU Advisor: Bing Luo
• Can phones train AI together without sharing private data? Authored the FedKit SDKs, the 1st practical way to
do this through real-world FL experiments across Android and iOS devices, which enables training shared ML models
collaboratively without sharing private data on phones.
◦
Streamlined research: automatically converted model formats and pushed models directly onto the phones.
• Wrote the official Android Kotlin example for Flower, the popular FL framework by a Y Combinator startup founded by
Cambridge University researchers. Consequently, got invited to speak at the Flower AI Summit.
◦
Helped correct the iOS example after finding that it was not training.
• Led a team of 6 to develop the FedCampus app, an Android & iOS app that leverages personal health data from over 100
participants to conduct real-world FL and federated analytics experiments on campus.
◦
Mentored the team, 4 undergrads and 1 M.Eng. student; interviewed and recruited 2 of whom.
◦
Investigated and led core technology adoption, including Flower, Kotlin, and Flutter.
Research Assistant, Search Engine Research Dec. 2021 – May 2023
The Search So Team, DKU Advisor: Jiang Long
• Developed a feature-rich open source web scraper in async Rust to scrape DKU and Duke intranet sites.
• Improved backend HTML processing, frontend interface, and version control.
Sichang “Steven” He | University of Southern California Page 2 of 3