Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Everyday tasks

T21: a research assistant that reads results first

Written November 12, 2026

A peer-reviewed Down syndrome literature assistant that prioritizes experimental results over abstracts. Why section-aware retrieval matters.

The problem with reading abstracts

Down syndrome, caused by trisomy 21, comes with a wide range of co-occurring conditions, and the research literature is large: nearly 35,000 publications indexed in PubMed as of late 2026, on the count given in the Scientific Reports paper by Sutanu Nandi and co-authors, published on November 10, 2026. A family member, a clinician, or a researcher trying to keep up faces the same wall. General-purpose language models answer from broad training data rather than from specific papers. Retrieval-augmented generation (RAG) improves on that by grounding answers in source texts — but a standard RAG system ranks background text as highly as experimental results, so an abstract's promise can carry the same weight as what a study actually found.

How the T21 Research Assistant retrieves

The T21 Research Assistant is a section-aware RAG system: it indexes papers by section and prioritizes Results sections when it answers, grounding responses in primary experimental evidence first. Its corpus is deliberately narrow — 1,594 open-access original research articles on Down syndrome from PubMed Central, including 158 studies funded by the NIH INCLUDE project. Around the retrieval sits a multistage pipeline covering query validation, retrieval, reranking, synthesis, and citation verification, built on open-source NVIDIA-Nemotron models. Answers come back structured and cited, so a reader can follow any claim to the paper behind it. The interface is accessible online, and the platform is open source, so it can also be deployed locally.

What the evaluation found

The authors tested the assistant on 30 expert-curated questions. The honest headline is parity, not victory: across most metrics, no statistically significant differences were detected between the T21 Research Assistant and leading proprietary and open-source models, or between it and two published biomedical RAG baselines. Where it did separate itself was against its own generator model used on its own — there, the full system significantly outperformed the standalone model on Recall, a measure of how much of the relevant material an answer covers. And its answers stayed traceable and anchored to citations, which is the property a reader of medical literature needs most when deciding whether to trust a summary.

Thirty questions is a small test. It is enough to show the design works as intended and holds its own; it is not enough to rank this assistant against every alternative a reader might try.

What it does, and what it does not do

Related playbooks

Sources

Back to: Everyday playbooks hub · Also: A human-agent team ran X-ray experiments you can audit · Find a DOI by title · AlphaGenome 2 reads variants per cell type · Luma: an NVDA add-on that describes the screen