Robin WU @RobinAIagent

I build AI systems for scientific research. Biomedical engineer at Oxford — agents, MCP servers, quality gates. I think by building, and I build what I think. robin.myaiagents.cc Oxford, England Joined December 2023

Tweets

123
Followers

16
Following

319
Likes

52

Robin WU @RobinAIagent

3 weeks ago

I’ve been building Robining Agent, a local-first CLI for developers and Agent builders. The idea is simple: treat an AI Agent as inspectable LEGO. Each task is classified as WHY, HOW, or MIX, routed through six semantic buckets, and reported with clear evidence, assumptions, and limitations. The runtime can also read and edit files, run approved shell commands, keep a conversation session alive, and switch between providers such as DeepSeek, Kimi, GLM, OpenAI-compatible endpoints, and Anthropic. The latest release is available through npm: npm install --global robining-agent robining I’m building this as an open-source framework for people who want Agents that are easier to inspect, extend, and trust. Feedback from developers and Agent builders is very welcome. GitHub: lnkd.in/eRjU_mwK #AIEngineering #OpenSource #AIAgents #TypeScript

1 0 0 19 0

View Details

Robin WU @RobinAIagent

4 weeks ago

Amazing!

Google Research @GoogleResearch

4 weeks ago

Introducing GlucoFM, a lightweight, self-supervised continuous glucose monitoring foundation model that separates metabolic baselines from transient spikes, producing transferable representations and setting new performance standards across diverse metabolic prediction tasks,

76 493 4K 363K 2K

0 0 2 18 0

View Details

Robin WU @RobinAIagent

a month ago

That is really cool!

Nous Research @NousResearch

a month ago

Introducing Bot Mode for Hermes Desktop. Your agent profiles become a series of named Bots. Each Bot has its own role, model, memory, skills and profile picture; Bots can use any model and even communicate with each other. Build a specialist Bot once to use it forever.

564 818 9K 2.5M 5K

0 0 1 20 0

View Details

Robin WU @RobinAIagent

a month ago

@ScienceMagazine Amazing ! The magical organization of cells!

0 0 0 112 0

View Details

Robin WU @RobinAIagent

a month ago

@ScienceMagazine Amazing

0 0 0 281 0

View Details

Robin WU @RobinAIagent

2 months ago

I’m excited to open-source Robin NMR Compound Inference Skill, an AI-powered capability for inferring candidate molecular structures directly from ¹H NMR spectra. Unlike a standalone AI agent, Robin is designed as a reusable scientific Skill that can be integrated into broader AI research workflows. It combines spectral interpretation with LLM-based reasoning to generate candidate molecular structures, SMILES representations, and confidence-ranked predictions. I hope this project can serve as a building block for AI-assisted chemistry, automated scientific discovery, and next-generation research workflows. 🔗 GitHub: github.com/Yaobin29/Robin… #AI #Chemistry #NMR #LLM #OpenSource #DrugDiscovery #Cheminformatics #ScientificAI #MachineLearning #Bioinformatics

0 1 2 136 0

View Details

Robin WU @RobinAIagent

2 months ago

@0xCarnagee Thanks for sharing this sososo amazing paper

0 0 0 74 0

View Details

Robin WU @RobinAIagent

2 months ago

@xie_yaqi @UofIllinois congrats!

0 0 0 4K 0

View Details

Robin WU @RobinAIagent

2 months ago

Beyond translated figures and single-paper summaries: I built research-gap-to-idea to compare papers, trace causal links, find research gaps, and turn open questions into falsifiable ideas. robin.myaiagents.cc/notes/why-i-bu…

0 0 2 20 0

View Details

Robin WU @RobinAIagent

2 months ago

@AndrewYNg So amazing! I am developing an agent about "from clinical data to organ on a chip". Would it be possible to have a chat with you? Thanks so much!

0 0 0 70 0

View Details

Robin WU @RobinAIagent

2 months ago

@JensenHuang @nvidia Totally agree that open source is sososo important!

0 0 0 3 0

View Details

Robin WU @RobinAIagent

2 months ago

MedPMC paper: 6.1M papers → 11M medical image-text pairs. 95.3% clinical relevance (vs 19.7%). MedPMC-CLIP: +7.1% AUC on 26 benchmarks with half the training data. Quality over quantity. arxiv.org/abs/2607.07673

0 0 0 12 0

View Details

Robin WU @RobinAIagent

2 months ago

🚀 AutoScientists: Self-Organizing AI Teams for Scientific Discovery What if AI could work like a real research team instead of a single assistant? AutoScientists, developed by researchers at Harvard University, introduces a decentralized multi-agent framework where AI scientists autonomously form teams around promising hypotheses, run experiments in parallel, share both successes and failures, and continuously adapt their research strategy—without relying on a central planner. The results are impressive: 🔹 BioML-Bench: 74.4% average leaderboard percentile, outperforming the previous best AI agent by +8.33%. 🔹 GPT Training Optimization: Reached target validation loss 1.9× faster, with 7 accepted improvements compared to 0 from a single-agent baseline. 🔹 ProteinGym: Improved protein fitness prediction by +6.5% on average across 217 DMS datasets, including a +12.5% gain on ACE2–Spike binding prediction. This work suggests that the future of AI for science may not be a single super-intelligent model, but self-organizing teams of specialized AI agents collaborating like human research groups. 📄 AutoScientists: Self-Organizing Agent Teams for Long-Running Scientific Experimentation 💻 Code: github.com/mims-harvard/A…

0 0 0 40 0

View Details

Robin WU @RobinAIagent

2 months ago

@bojie_li @nsdevaraj @toanalien So amazing to see such brilliant book sharing in github!

0 0 0 129 0

View Details

Robin WU @RobinAIagent

2 months ago

Built by fine-tuning DrugGPT (GPT-2) with: ① Supervised fine-tuning on 13,908 disease-target-drug triples ② GRPO reinforcement learning (same technique as DeepSeekMath) ③ 3 reward functions: binding affinity (PLAPT), novelty, diversity Paper: arxiv.org/abs/2607.08404 Code: github.com/alimotahharyni… 🧬💊

0 0 0 12 0

View Details

Robin WU @RobinAIagent

2 months ago

DrugGen-2: A disease-aware GPT-2 for drug discovery Most AI drug design models only look at the target protein. But the same target behaves differently in different diseases. DrugGen-2 conditions on BOTH disease ontology (MeSH) AND protein sequence. Results are striking: 🧵

2 0 0 20 0

View Details

Robin WU @RobinAIagent

2 months ago

Key results: • 409-444 unique molecules generated per target (vs DrugGen: 50, DrugGPT: 219) • 99-100% chemical validity • Binding affinity scores: 9.26-9.97 median vs baselines at 5.86-8.49 • Docked ligands beat reference drug enalapril for ACE target (-9.917 vs -8.283) All on 5 diabetic nephropathy targets.