Savviety

Work

My software career and AI work.

I bring what I've learned as an engineer, architect, and delivery lead to my AI projects. The examples below explain the decisions I made and link to the code and records so you can check the work.

Before the lab

I've worked in six startups through an operating career that included four exits. Since 2016, I've run Savviety as a fractional CTO practice. The systems I've built, architecture decisions I've made, and teams I've led inform my work at Savviety AI Lab.

  • Interclick

    I built and scaled the ad-targeting platform behind a $270M exit.

  • Kroger

    I led Kroger's first cloud implementation.

  • Magic Leap

    I held an engineering leadership role during the company's build-out.

  • Applied ML

    My work includes machine learning in production for audiology and ML delivery for a Big Four firm.

Decisions in three projects

Software engineering · savviety-instinct

Testing the assumptions behind the code

A code-analysis tool needs accurate measurements. If the tests and implementation share the same assumptions, both can miss the same bug.

I checked the algorithms against independent implementations, property-based invariants, and adversarial parser fixtures. That's how I approach code written with AI: decide what correct means, then try to find where the code gets it wrong.

These tests found three bugs the original expectations missed. One external checker had the same blind spot as my code; a second checker disagreed and exposed it. The repo documents the defects and the checks that caught them.

Read savviety-instinct on GitHub

Architecture · kb

Keeping note capture working when the server is down

I need to be able to save notes when the search server or model service is down. Capturing a note shouldn't depend on the services needed to retrieve it later.

The capture commands write Markdown locally. A separate server indexes the notes in Qdrant, embeds and reranks them with Voyage AI, and answers questions with citations through an OpenAI-compatible model server such as oMLX. Saving a note doesn't need a network connection.

The tool works as a personal knowledge base today. The repo includes dependency health checks and a smoke-test runbook. MCP integration is documented as future work.

Read kb on GitHub

Delivery · opening-dojo

Building a chess trainer with AI agents

The chess trainer needed to work as a coherent application, with tests to check its behavior. To coordinate the agents, I also needed a clear scope and a definition of what each task had to deliver.

I designed the process and directed the work, starting with requirements and a plan. Agents wrote the implementation, which went through several rounds of review. I set the direction and decided what had to be true before I would accept the result.

The repo has 86 tests in CI. It also includes the product requirements document (PRD), execution plan, review records, links between requirements and tests, and a postmortem. You can read the delivery records alongside the code.

Read opening-dojo on GitHub

Public repos

These repos include applications, engineering tools, and experiments. Each documents what it does and how it's checked. The projects built with AI show how I work today, following the earlier career work above.

Verifiable delivery

Modernization and behavior

Verification and analysis

Retrieval and knowledge

Applications

Talk through your project