Two things we would like to research with someone.
Concrete enough to write a bid around. We bring running agent systems, evaluation tooling and the engineering; a partner brings the problem, the data or the discipline we lack.
- 01
Isolation ladders for agent tool use
Our Serang work runs agent-written code through a ladder of sandboxes (isolates, containers, gVisor) with per-run budgets and an egress proxy. The open question is cost-effective assurance: which rung is enough for which class of task, measured on real agent workloads rather than synthetic benchmarks.
Who we are looking for: A security or systems group, or an organisation running agents on sensitive data.
- 02
Claim verification at reading speed
Sagan extracts structured claims from papers, compares them across papers and, where code is public, reruns the experiment. We want to measure how often verification changes a reader's conclusion, and what a verification record should contain to be trusted by a research team.
Who we are looking for: A university library, a research group or a journal interested in reproducibility.
How it works
We co-write the proposal, share the engineering openly where the funder allows it, and publish what we learn as field notes. Student projects and placements are welcome. We are based in the UK and work with Innovate UK, UKRI and EU formats.