Research
Published work below, and — more usefully — the open problems. Those are listed with what we are missing, because the fastest way to find a collaborator is to say plainly what you need.
Published
arXiv preprint · 2025
Agents that call databases, shells, and web APIs through Model Context Protocol inherit every permission those tools carry. This paper proposes a security layer that sits in front of MCP traffic — authentication, rate limiting, logging, and scanning — so an agent's blast radius is something you configure rather than discover.
Read the paper ↗Open — collaborators wanted
None of these are finished. Each one lists what would move it forward — if that describes you, the contact form is enough.
01
Can a model read a COBOL or legacy Java estate well enough to make migration a tractable engineering project rather than a multi-year rewrite?
Enterprises hold decades of code nobody currently understands. The interesting part is not translation — it is establishing behavioural equivalence you can defend to an auditor.
Looking for
02
How should an agent choose between hundreds of available MCP tools without putting all of them in the context window?
Tool count is growing faster than context budgets. We are building modular routing infrastructure that selects, ranks, and scopes tools per request, and measuring what that costs in accuracy.
Looking for
03
Where does symbolic reasoning beat a language model outright, and can the two be combined without inheriting the weaknesses of both?
Regulated sectors need decisions that can be explained after the fact. Pairing learned components with symbolic constraints is the most promising route we have found, and the least well measured.
Looking for
FAQ
Agent security and observability, mostly — the parts of agentic systems that only become visible once something is deployed. Published work so far covers Model Context Protocol security; the open problems extend into tool routing, legacy code understanding, and neuro-symbolic reasoning.
Send a note through the contact form saying which problem and what you would bring — a dataset, a domain, an implementation, or time. There is no application process and no institutional affiliation required. What matters is that you will actually run experiments.
No. Several of these problems need practitioners more than researchers — someone running fifty tools against an agent in production knows things no paper currently records.
Not through grants. It is done alongside the engineering work, and GitHub Sponsors covers costs like publication fees and compute. That means timelines are honest rather than promised.
Preprints go to arXiv and are linked here on publication. Code goes to GitHub under the same open licences as everything else. Google Scholar has the full citation record.
Yes, and it happens. Get in touch first so the scope can be cut to something a semester actually fits, and so you are not duplicating work already in progress.