Practice

Two team leaders coached their colleagues on schedule adherence through brief, informal check-ins. I recognized their contribution publicly.

2026

The work combined program management with engineering. I developed the replacement CRM and its data model while coordinating with the operations, engineering, and data teams involved in the transition.

I documented field definitions and lineage in a data dictionary spanning at least 20 databases and other data sources in total, accessed through the lakehouse and directly. We joined account information with product usage, telemetry, and customer input in complex queries exposed through callable API endpoints.

August to October 2026

I worked on a knowledge base with vector retrieval so a copilot could draw on the same underlying content. I also implemented an agentic documentation pipeline that scanned engineering repositories, identified code changes, and drafted updated knowledge-base articles for human review before publication.

The draft-to-publication boundary mattered. Article generation ran regularly, while people remained responsible for reviewing and publishing the resulting content.

2025 to 2026

As Director, I worked on vendor coaching prompts that accounted for sentiment, conversation stage, and a quality rubric. Later work across my Nerdy tenure included an evaluation workflow with iterated judges and alerts for leaders. I built the evaluation workflow with open-source components.

The intended use was concrete: point a manager toward a conversation worth reviewing, whether it offered a chance to recognize good work or coach a difficult interaction. The manager still needed to read the conversation and use their judgment.

2024 to 2025

The implementation sat within my Marketplace Operations role. Alongside the platform work, I partnered with Finance on headcount forecasts and resource allocation.

That work connected to the broader operating role: bringing tutor engagement and support functions together, introducing Agile and Kaizen practices, and creating mentoring opportunities for leaders.

2022 to 2024

I developed segment-based measures that distinguished agent and customer response timing. That made the definition of a measure part of the work, alongside the dashboard used to display it.

I also partnered with AI/ML and quality teams on scoring rubrics and calibration, and contributed to the implementation of a coaching platform. These contributions belonged to my Chat operations role, before the separate Program Manager assignment.

2013 through 2017

I designed the stationery and programmed laser-guided die cutters and four-color process printers. I built the CRM from scratch, along with the online store, customer service platform, and shipping integrations. Marketing and social media were part of the work too.

I also handled the business license, bookkeeping, and accounting.

A useful AI assistant needs to understand the job it is being asked to do. It needs enough context to ask a relevant question, produce a usable draft, and leave the person using it able to decide what happens next. Much of my work has been about making that setup easier to repeat.

Turning an idea into a project

At Nerdy, I created project-building and prompt-engineering resources that helped colleagues describe a task and assemble the instructions and context for an assistant. I taught the approach so people could adapt it to their own work.

Organizing the working day

I also developed and taught a reusable work-organization assistant. It helped bring information into a form a person could review, prioritize, and act on. A colleague's reuse of the approach was one useful sign that the pattern could travel beyond my own setup.

The emphasis was on giving the person a clearer working view. They still made the decisions, checked what was missing, and chose which actions to take.

Helping a request reach engineering

Another workflow guided colleagues through describing a request for an engineering ticket. It brought the problem and relevant context into the conversation so the request could be made more specific before reaching the team.

That experience informs how I approach product development: understand the person and the task, make the context available, and decide where the system should ask for a human decision. I bring the same approach to custom applications, retrieval, plugins, and MCP integrations in my independent work.

AI enablement begins with the work someone already understands. A manager has a coaching conversation to prepare for. A colleague has a project they want to start. The useful question is what would help them make progress, and how they will know the result is worth using.

Coaching and operational judgment

At Chewy, I built a scorecard to keep goals, previous conversations, and employee comments together. I taught managers how to use ChatGPT in their coaching work and later built a custom GPT around that workflow. Eleven chat managers used the approach while supporting roughly 180 to 200 agents.

The early process used manual handoffs. Managers supplied the material, reviewed the output, and remained responsible for the coaching conversation. I also managed AmplifAI onboarding for the chat team and trained participants in using the platform.

Alongside that work, I used SQL, Vertica, Excel, and Tableau to investigate operating questions. I checked analysis with fellow managers and helped Operations HR compare employee reviews during calibration. The people responsible for those decisions stayed involved.

Reference material and live help

At Nerdy, I wrote a 263-page AI guide and held office hours where colleagues could bring questions and work through their own projects. The written material gave them something to return to after a session. Live help made room for the question that only becomes clear when someone tries the work.

I developed reusable project patterns and taught colleagues how to adapt them. My manager recognized the guide as an organizational resource and the teaching as practical support for peers moving from a blank repository toward a first deployment.

These experiences shape my consulting work. I can help a team choose a useful starting point, build supporting resources, and work through the decisions that arise as people begin to use AI.

I converted and published six Hemmingway-1 builds for MLX and oMLX on Apple Silicon, with model cards that describe the source revision, quantization choices, calibration, and verification. Altworld developed the source model. My work concerns making those weights usable locally and examining the tradeoffs.

The collection spans oQ2e through oQ8e. Its companion preference study compares a fixed writing suite with a local BF16 reference. A separate evidence package examines BF16-to-oQ4e numerical fidelity and records controlled oQ4e runtime behavior. Each answers a different question, and the published limits keep those conclusions separate.

The public record includes a correction to the first aggregate analysis. I want someone reading the work to be able to see what changed, inspect the supporting material, and understand what still needs to be tested.

Explore the model cards, studies, and correction record through the evidence links below.

I build applications and integrations that give AI tools useful context and a clear job to do. This site's editorial setup connects a browser-based content system with MCP tools for finding and editing records. Drafts stay separate from public content, and access is scoped to the editorial work.

That work sits alongside my interest in retrieval and local models. I want the connections to be understandable: where information came from, what a tool may change, and how a person can review the result.