The knowledge was already there. The system to use it was not.
The support problem that pushed me toward applied AI was not a lack of answers. It was repeatedly rediscovering knowledge the organization had already paid to discover.
Writing
Engineering write-ups on building, evaluating, debugging, and operating AI systems. Local essays appear alongside selected work from the Keystone blog.
Published
Regulated AI has a retrieval problem, not a knowledge problem
A bounded engineering case study in hybrid retrieval, authorization eligibility, evidence thresholds, source-oriented answers, and explicit refusal behavior.
What evaluation artifacts taught me that demos never could
Why retaining failures, binding results to code and configuration, and separating strict assertions from characterization changed how I evaluate AI systems.
From Genesys to Keystone: enterprise rigor on the LLM substrate
What 12+ years in enterprise contact-center systems carried into hands-on AI engineering, and which probabilistic, retrieval, agent, and evaluation problems required new methods.
The Six Substrate Dimensions
A proposed test for what might belong in the substrate: if you replace the agent, framework, or provider and the property is still necessary, it may be a substrate candidate. The piece examines six candidate dimensions as research hypotheses rather than an established ontology.
The Protocol Between Agents Is the Protocol Between Teams
Reflections from 12+ years working with enterprise contact-center systems, and a hypothesis about how interface changes, organizational boundaries, and accumulated governance mechanisms may reappear in multi-agent AI.
Causal Decomposition Under Suspended Judgment: Two Cases from Keystone
Two cases in diagnosing governed AI failures at the mechanism level. A retrieval system returned wrong-topic chunks with high confidence; the actual causes were three independent factors, none of them the first hypothesis. The second case applies the same decomposition discipline to an architectural decision made before the system existed.
Provana AcuteCare: clinical copilot with guardrails
A clinical copilot with role-based action guardrails: three protocols, generative UI, four governance outcomes (Serve, Block, Route, Mutate). The prototype reused the same governance architecture from workplace safety to acute-care medicine without structural changes in this implementation, an engineering observation from one prototype, not evidence of general portability. Built at the Boston AI Tinkerers hackathon with Sai Gopal Jarabana.
Governed incident response: agentic UI with RBAC and audit trails
An agentic incident-response prototype where role checks gate tool calls, actions are written to a hash-chained audit record, and the generated UI reflects the actions available to the user's role. Solo build for the AI Tinkerers Generative UI Hackathon.