A research assistant that reads the literature overnight
Confidential · Social Impact
The business said
- “Hire more analysts to read papers”
- “Subscribe to yet another citation database”
- “Commission a consultant study per intervention”
- “Build a document repository with a search box”
- “Let a researcher test a theory of change against the world's evidence in near real time”
Assessing an early-stage poverty intervention was not a reading problem. It was a retrieval and mapping problem: find the evidence, arrange it into an impact pathway, and keep the researcher in charge.
A generative AI research assistant, delivered under a cloud research grant
A global nonprofit fights poverty by backing innovations for people who feed and educate their families on a few dollars a day. Before committing to an intervention, its research team must check the idea against the published evidence: does a clear pathway from action to impact actually exist? Doing that by hand meant weeks of searching scholarly databases, development-bank archives, and the open web, then reading and sorting whatever came back. The asks all circled the same pain: more readers, more databases, more consultants.
We built an automated research assistant instead. A researcher types a query and the system fans out across sources such as Google Scholar, the World Bank, and general search engines at once, with workflow orchestration triggering the retrievals in parallel. Everything found lands in a personal data library with the title, summary, source, and authors of each document. Libraries can be edited, merged, and shared between researchers, and a question-and-answer interface lets the researcher interrogate the collection directly.
The assistant then drafts the thinking, not just the pile. The researcher picks a preferred large language model from the major providers, and the system produces an impact-pathway outline of topics and subtopics, tagging every retrieved document to the branch it supports. The researcher stays the expert: they can prune or add subtopics, delete or add documents, fire follow-up sub-queries that pull a fresh wave of evidence, and regenerate the outline with retrieval-augmented generation. Graph storage keeps the pathways connected and explorable.
The platform runs on managed Kubernetes in the cloud, with pipeline orchestration for every automated step, and was delivered under a competitive research grant from the cloud provider. What used to be weeks of desk research became a same-day loop: query, evidence library, impact pathway, decision. The nonprofit gets what philanthropy rarely has, which is accountability for results before the money moves.
- cloud research grant, funded by
- scholar, dev banks, web, evidence swept
- 3 major providers, LLM choice
- editable + shareable, libraries
- weeks to same day, desk research
