I build autonomous research agents that improve their own code, tools, context and skills.
Note
AI-generated optimization PRs: This account is used to submit optimization pull requests written by autonomous AI agents as part of our research at Weco AI. We're testing how well agents can identify performance improvements, implement and evaluate changes, and prepare PRs for real open-source projects.
I'm Dex, a Member of Technical Staff at Weco AI in London, UK. My agents write code, test it against a benchmark and keep what scores better. I co-authored AIDE, the tree-search ML agent OpenAI used in MLE-bench, and AIDE², which runs that loop on AIDE's own code. I also built the research agent that set seven records in OpenAI's Parameter Golf.
I've contributed to AIDE since 2024, and our 2025 paper describes how it works. Meta FAIR reimplemented AIDE as a baseline for its own research agents.
In AIDE², an outer agent rewrites AIDE's own code and keeps a rewrite only when it scores higher on held-out tasks. In an eight-day unattended run it found seven successive improvements, and the strongest discovered agent matches or beats the hand-tuned production agent on four held-out benchmarks: ALE-Bench, MLE-Bench, FML-Bench and WeatherBench 2. On a separate held-out task family, the discovered agents' reward-hacking rate fell from 55% to 32%, below the hand-tuned agent's 39%. The details are in our technical report, and in the Weco AI post we place this at Level 1, "net positive", on a four-level ladder of recursive self-improvement.
Previously, I built trading systems at Hex Trust. I also served as a maintainer of the Hyperledger Fabric Python SDK, a Linux Foundation project.
I studied Information and Computing Science at the University of Liverpool and Xi'an Jiaotong-Liverpool University. I also did research at Nanyang Technological University, Zhejiang University, and Hong Kong Baptist University.
- AIDE — contributions to the open-source agent that writes, evaluates, and improves machine learning code. I also contribute to weco-cli, the command line tool that drives it.
- OpenAI Parameter Golf — my research agent set leaderboard records in 2026, with entries in the official record-track directory. The agent's best result achieved a 5-seed mean validation BPB of 1.0645. OpenAI also highlighted one of the agent's model-compression PRs in its write-up about the competition.
- UK AI Security Institute's Inspect — performance work on the UK government's LLM evaluation framework: cut clustered-stderr scoring time and memory (#4714) and made tool-result media extraction linear in conversation length (#4628).
- Agent runtimes — merged performance work into openclaw, AutoGPT, goose, qwen-code, pydantic-ai, agno and BAML.
- AI research infrastructure — cut redundant AST parsing in Sakana AI's ShinkaEvolve and bounded process-pool shutdown latency in OpenEvolve; smaller docs fixes in Meta's aira-dojo, Microsoft's RD-Agent and OpenAI's MLE-bench.
Each project links to its merged pull requests on GitHub. Ranked by stars, refreshed weekly.
Contributions to 130 open source projects — 59 of them AI or agent infrastructure.
15 more AI and agent infrastructure projects (under 1,000 stars)
| Project | Stars | Contributions | Latest |
|---|---|---|---|
| 907 | View PRs | Nov 2026 | |
| 892 | View PRs | Nov 2026 | |
| 829 | View PRs | Nov 2026 | |
| 805 | View PRs | Oct 2026 | |
| 802 | View PRs | Nov 2026 | |
| 689 | View PRs | Oct 2026 | |
| 639 | View PRs | Oct 2026 | |
| 602 | View PRs | Nov 2026 | |
| 578 | View PRs | Nov 2026 | |
| 577 | View PRs | Oct 2026 | |
| 440 | View PRs | Mar 2026 | |
| 179 | View PRs | Jul 2025 | |
| 124 | View PRs | Sep 2025 | |
| 35 | View PRs | Jul 2026 | |
| 3 | View PRs | Aug 2026 |
32 more non-AI projects (under 1,000 stars)
- Recursive Self-Improvement of AI Research Agents (arXiv), arXiv preprint, 2026
- AIDE: AI-Driven Exploration in the Space of Code (arXiv), arXiv preprint, 2025
- Lightweight and Unobtrusive Data Obfuscation at IoT Edge for Remote Inference (DOI), IEEE Internet of Things Journal, 2020
- Challenges of Privacy-Preserving Machine Learning in IoT (DOI), ACM AIChallengeIoT, 2019
- A Deep Reinforcement Learning Framework for the Financial Portfolio Management Problem (arXiv), arXiv preprint, 2017
- Hands-on AutoResearch: Cracking OpenAI's Parameter Golf — workshop with the Weco AI team, AI Engineer World's Fair 2026
- Algorithmic Trading Workshop — Network School, first cohort (2024)
- Deep Learning for Power System Security Assessment (2019)
- Introduction to Hyperledger Fabric (2019)
- 🏆 Dev Bounty Winner, Network School, first cohort (2024)
- 🏆 Special Prize (US$10,000), Wanxiang Blockchain Hackathon by Qtum (2018)
- 🥇 1st Prize, EOS Hackathon Hangzhou (team, 2018)
- 🥇 1st Prize, Hack x FDU 2017 Hackathon (out of more than 70 teams)
- 🥈 2nd Prize, XJTLU Blockchain Technology Application Innovation & Entrepreneurship Challenge (2020)
- 🥈 2nd Prize, XJTLU & PNP AI Innovation Hackathon (2018)
- 🥉 3rd Prize, EOS Hackathon Hangzhou (individual, 2018)
- 🥉 3rd Prize, DoraHacks x BCH Faith Hack (2018)
- 🏆 IBM Student Innovation Lab Program Award (2018)
- 🎓 Hyperledger Diversity Scholarship, Hyperledger Global Forum 2020 (2019)
- 🎓 CNCF Diversity Scholarship, KubeCon + CloudNativeCon China (2018)
An open-source tool I built: WakaTime-style usage tracking for Claude Code, Codex, and OpenCode. The charts below show my own usage, with repeating animations and daily updates.
