How I built an AI platform for R&D teams

The Science Room page of the Hammer R&D platform demo

At Hammer, I'm building an AI platform for R&D teams. Longer write-up is on the Hammer blog, read the more technical version with here.

The problem and the platform

Research teams build up years of experiment reports, measurements and papers. When a new idea comes up, the question is often whether someone tested it before. The answer is usually somewhere in a report, a table or a colleague who already left. Finding it takes a lot of time. We try to make this process way quicker and easier for the users to answer by building this AI-based R&D platform with their custom needs in mind.

The problem described above is actually part of a bigger, structural, problem which is grounded in academia. I will cover this later!

What I built

The platform is built with the analogy of a house. This house is guarded and built in the EU, so it falls under its strict protections. Each room in the house has its own functionality focus and is designed meticulously to accomodate that function. There's also a large secure safe in the house which stores the most valuable data of the company who uses it.

The various rooms (functionalities) are described below.

Science Room: ask questions across internal reports and uploaded papers, with references back to the source. Also works on tables, charts and images.
Paper recommendations: suggests related papers based on the topics in the collection, with a reason why.
Experiment review: splits a proposed experiment into claims and checks them against the literature for support, contradictions and gaps.
Patent screening: a first check for possibly overlapping patents.
Sharing: conversations and results can be shared within a team, with group permissions.

Answers separate what the documents say from what is a hypothesis. Everything runs on European infrastructure with Mistral models.

There's an interactive demo with fictional documents.

Technical hurdles and features we're proud of

Now that we know what we built, I want to share some more technical details on how we achieved it.

The extraction of information from the information-rich data in the vault is key to even begin having a useful and workable product...