Turning academic literature into research leads

A SaaS B2B business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas, each one a test of whether generative AI could be trusted in a rigorous scientific setting.

We structured the engagement in three phases so that each layer of capability could be built and tested before the next was added.

The proof of concept met its goal. It demonstrated that researchers could ask scientific questions in natural language and receive answers grounded in stored academic literature, with the reasoning traceable back to specific papers. In the scenario walkthrough, the agent

• Grounded answering proven — Researchers received scientific answers drawn from the ingested literature, with output explanations traceable to specific reference papers.
• Cross-domain reasoning demonstrated — The system linked genetic sequences to protein products, and protein functions to chemical pathways, across public resources.
• Agentic tool use validated — The workflow ran BLAST-like reasoning through the NCBI website and returned the results within a single answer.
• Handover — We documented and reviewed the LLM responses from the scenario walkthrough, leaving the client with a clear, evidenced view of what the approach can and cannot yet do.

Took a real molecular structure and reasoned through motif recognition, comparable metabolites and candidate biosynthetic gene clusters, exactly the chain of thought the engagement set out to validate.

With viability established, the natural next steps (projected) would be to broaden the ingested literature, harden the ingestion pipeline and expand the agentic tool set, moving the proof of concept towards a research assistant the client's scientists could rely on day to day.