Research
Scope
My research concerns systems that must remain understandable when state is distributed, connectivity is unreliable, or automated decisions are difficult to inspect. The work is practical: each area is tied to software I am building.
Local-first state and recovery
I study how an application can remain useful offline, merge delayed changes without hiding conflicts, and preserve an operator-readable history.
The main tools are CRDTs, append-only event records, deterministic identifiers, and recovery paths that do not depend on a service remaining available.
Private state and selective verification
Zero-knowledge proofs and private settlement are useful when a system needs to establish that a condition was met without exposing the complete underlying state.
Current work considers payment linkage, selective disclosure, and audit records whose claims can be checked without turning privacy into opacity.
Bounded context for LLM systems
A model's behaviour depends on what enters context, what is excluded, and how memory is compressed.
I am exploring constraint design, retrieval boundaries, provenance, and controllable memory for workflows where unsupported output is more costly than an explicit refusal.
Verification under changing evidence
Public analyses and operational systems need to absorb new evidence without rewriting history. I am interested in revision-aware records, confidence changes, and review queues that separate proposed evidence from published claims.
Method
The common method is to define invariants first, expose uncertainty, preserve source evidence, and test degraded modes rather than only the happy path.