Now
Working on
The agent harness behind OnFinance's enterprise deployments, and the deployments themselves: multi-agent workflows for regulatory compliance audits, financial risk modelling, and accounting operations, running inside banks' own environments.
Thinking about
- How to make an evaluation suite strict enough to block a bad release but cheap enough that nobody is tempted to skip it.
- What an agent's output needs to carry with it so that a risk team can audit it without reading a trace.
- Where deterministic code should end and the model should begin, inside a harness that has to run on-prem.