Optimizingtheoptimizer.

I build recursively self-improving AI agents.

I'm Dex, a Member of Technical Staff at Weco AI (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.

See selected work → GitHub profile

PathBlockchain infrastructure, then distributed ML, now self-improving research agents

2016-18XJTLUResearch assistant; deep RL for portfolio management
2018-24HyperledgerIntern, then Fabric Python SDK maintainer
2019-20NTU · Zhejiang U.Private ML at the IoT edge
2020-21HKBUAll-reduce for distributed training
2022-24Hex TrustTrading backend in Rust and Go
2024-nowWeco AIAI research consultant, then Member of Technical Staff

Languages: PythonRustGoTypeScriptC
ML & agents: PyTorchLLM agentstree searchevals
Infrastructure: KubernetesTerraformNginxHorovod
Agent tooling: HarborLangChainDeep AgentsPi

Proof

Numbers anyone can check on GitHub.

open-source projects with my merged code

180⁺

merged pull requests

records in OpenAI Parameter Golf

4.3k

stars on seedance2-skill

Cited, used or merged by

OpenAI

Used AIDE for MLE-bench in the GPT-4.5 system card, and featured my agent's PR in the Parameter Golf write-up.

System card ↗Write-up ↗PR #1060 ↗

Meta FAIR

Rebuilt AIDE as a baseline in AIRA-dojo.

arXiv 2507.02554 ↗

Google Cloud

Benchmarked its MLE-STAR agent against AIDE.

arXiv 2506.15692 ↗

Sakana AI

Cited AIDE's tree search over code in The AI Scientist-v2.

arXiv 2504.08066 ↗

METR

Ran the AIDE scaffold in RE-Bench, its AI R&D benchmark.

arXiv 2411.15114 ↗

UK AI Security Institute

Merged my speedups into Inspect, its LLM eval framework, in 2026: faster clustered-stderr scoring and linear-time media extraction.

inspect_ai #4714 ↗inspect_ai #4628 ↗

Selected work

Agents that keep working
after I close the laptop.

Weco AI · co-author · 2026

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.

first evidence of recursive self-improvement
reward hacking 55% → 32%

Read the post Technical report

Weco AI · co-author, co-maintainer

AIDE

Tree search over code: draft, run, debug, improve, keep the best.

1.6k stars
top scaffold in OpenAI's MLE-bench paper

Read the paper Code

OpenAI challenge · 2026

Seven records in Parameter Golf

Best language model in 16 MB and 10 minutes of training. My research agent ran for 22 days, filed 25 PRs of its own and set twice as many records as the best human, out of 1,016 entrants.

cited 435 times by other entrants

Record track Best entry

Agent skill · 2026

seedance2-skill

Turns a rough idea into a production prompt for the Seedance 2.0 video model.

4.3k stars · 347 forks

Repo Site

Developer tool · 2026

Vibe Clock

WakaTime for coding agents: tracks Claude Code, Codex and OpenCode, renders SVG cards for your profile.

Python · SVG · Homebrew

Repo

Linux Foundation · 2018-24

Fabric Python SDK

Maintainer of Hyperledger's Python SDK. Ran the v0.7 release, mentored two internships and interned on Cello.

Maintainer · release manager

Maintainers file

Open source

Approved by maintainers.

Speed and correctness fixes to other people's projects. Many recent ones were written by the agents I build at Weco AI, and each went through the maintainers' normal review.

openclawopenclaw★ 392k

AutoGPTSignificant-Gravitas★ 188k

langchainlangchain-ai★ 148k

OpenBBopenbq-org★ 74k

remotionremotion-dev★ 63k

gooseaaif-goose★ 55k

agnoagno-agi★ 43k

manimManimCommunity★ 41k

pnpmpnpm★ 37k

poetrypython-poetry★ 34k

InvokeAIinvoke-ai★ 29k

qwen-codeQwenLM★ 28k

llamafilemozilla-ai★ 26k

backtradermementum★ 23k

oxcoxc-project★ 23k

pydantic-aipydantic★ 21k

16 of 59 projects shown, by stars. Full list on my GitHub profile.

Highlights

2026
Seven agent runtimes
Performance fixes in openclaw, AutoGPT, goose, Qwen Code, Pydantic AI, agno and BAML.
2026
Evolutionary search
Cut redundant AST parsing in Sakana's ShinkaEvolve and bounded shutdown latency in OpenEvolve.

Papers, talks, awards

Papers and talks.

Publications

2026 · arXiv · AIDE²Recursive Self-Improvement of AI Research AgentsarXiv 2609.26457
2025 · arXivAIDE: AI-Driven Exploration in the Space of CodearXiv 2502.13138
2020 · IEEE Internet of Things Journal · first authorLightweight and Unobtrusive Data Obfuscation at IoT Edge for Remote InferenceDOI 10.1109/JIOT.2020.2983278
2019 · ACM AIChallengeIoTChallenges of Privacy-Preserving Machine Learning in IoTDOI 10.1145/3363347.3363357
2017 · arXivA Deep Reinforcement Learning Framework for the Financial Portfolio Management ProblemarXiv 1706.10059

Talks

2026 · AI Engineer World's Fair, San FranciscoHands-on AutoResearch: Cracking OpenAI's Parameter GolfWorkshop with the Weco AI team
2024 · Network School, first cohortAlgorithmic Trading WorkshopSlides
2019Deep Learning for Power System Security AssessmentSlides
2019Introduction to Hyperledger FabricSlides

Selected awards

Dev bounty winner, Network School (Farcaster social-graph visualiser), 2024
Special Prize (US$10,000), Wanxiang Blockchain Hackathon by Qtum, 2018
1st Prize, EOS Hackathon Hangzhou (team), 2018
1st Prize of 70+ teams, Hack x FDU, 2017

Get in touch

Felix, qui potuit rerum cognoscere causas.

Lucky is the one who has learned the causes of things. Virgil. It's also my GitHub bio.

Write to me about self-improving agents, evals, research infrastructure or an open-source project you want faster.

i@dex.moe Read the full CV