How LLMs and agent harnesses co-evolved from fixed workflows into deep research systems built around long-horizon search, coding agents, heterogeneous knowledge sources, and self-improvement.
How we built BioDSBench to evaluate LLMs on biomedical data science, found that analysis plans are key to success, and designed the DSWizard agent framework. From benchmark to platform to biomarker discovery.
LEADS: a foundation model for human–AI collaboration in medical literature mining (Nature Communications, 2025). Why proprietary data and domain workflow matter for systematic review—and how LEADS performs.
Using our BioDSA agent framework, we reproduced eight biomedical AI agents in one day. The real bottleneck wasn't intelligence — it was architecture friction. Here's what we learned about designing for rapid agent prototyping.
Generic Deep Research agents browse the web, but biomedical discovery requires navigating heterogeneous Knowledge Graphs. DeepEvidence is a specialized investigator that systematically explores biomedical KGs, achieving 40% accuracy on HLE-Medicine while frontier models hit only 3.3%.
Agentic AI systems like ChatGPT DeepResearch can browse the web, but biomedical literature is different. We care about high-quality evidence from published studies, clinical trials, and curated databases. Here's how we trained LLMs to become domain-specific literature search agents using SFT and RL.
Systematic literature reviews in medicine are more than just searching and summarizing papers—they're regulated processes requiring reproducibility and rigor. Learn how TrialMind-SLR turns PRISMA guidelines into an AI-powered workflow, and how the LEADS foundation model outperforms GPT-4o on clinical literature mining tasks despite being much smaller.
Generic AI models struggle with highly specialized domains like clinical trials—they don't know what to retrieve, improvise details that should never be improvised, and forget that regulators exist. Learn how we built AI systems for clinical trial document generation through retrieval, fine-tuning, and agentic workflows, from Trial2Vec to InformGen.
Predicting whether a clinical trial will succeed sounds like science fiction. Discover how machine learning models like HINT and SPOT are finding patterns in past trials to forecast future outcomes, and how agentic AI systems are transforming clinical development.
Patient recruitment is the headache no one talks about at parties. Trials stall. Budgets burn. Great ideas never meet the right people. Learn how TrialGPT uses generative AI to match patients with clinical trials.