Working papers

Writing

Two papers on building AI systems that reason over real data and can be trusted to act on it. The first is about the flow of a single conversation. The second is about the anatomy of the system that runs it in production.

Paper 1 of 2·On the flow of a single conversation.

Conversational Decision Intelligence

Flow, Reasoning, and Continuous Learning in Natural Language Systems over Structured and Unstructured Data

Organizations increasingly expect to ask questions of their data in plain language and receive accurate, explainable answers, whether the underlying data is structured (relational databases, tabular files) or unstructured (documents, application programming interfaces). Achieving this capability reliably requires more than a single language model invocation. It demands a conversation flow that plans, reasons step by step, recovers from failure, and improves with use.

March 202619 min readPublic, for practitionersRead the paper →

Paper 2 of 2·On the anatomy of the system that runs it.

One Agent, Many Hats

Rules, Skills, and Tools as the Architecture of an Autonomous, Self-Extending Agent System

This paper presents the architecture of the agent system built at Insighter: a single reasoning engine specialised at runtime by three declarative layers. Skills are versioned playbooks that define how the agent works toward a given outcome and allow it to adopt distinct behavioural roles (planner, evaluator, coder, guardian, critic, reflector, communicator) without spawning separate agents. Rules are guardrails enforced at three escalating strengths: as prompt instruction, as coded gates that block progress, and as hard infrastructure boundaries that the model cannot override. Tools are the only actions the agent can take: read-only, tenant-scoped, allowlisted per skill.

July 202637 min readRead the paper →