AI

We build agentic AI systems for investment research, trading, and football analytics. Our next generation is designed for long-running workflows, with persistent state, checkpointed execution, and typed tool interfaces. Agents will plan within explicit compute budgets, revise their plans as new evidence arrives, and invoke tools under scoped permissions, with human approval at defined decision points.

The architecture will combine vector retrieval for unstructured research with graph-based memory for relationships among entities, evidence, and decisions. We’re exploring Monte Carlo tree search for decisions with multiple viable paths and reinforcement learning from feedback scored against domain-specific rubrics. Evaluation will measure factual accuracy, tool use, and policy adherence across multi-step workflows.

We plan to record model versions, evidence references, tool calls, and approval events in a queryable audit trail, with cryptographic hashes to verify record integrity. Operators will be able to resume interrupted workflows and trace consequential decisions back to the data, assumptions, and actions that produced them.