Work

Selected work

These are selected projects, grouped by the topic they belong to.

Research and notebooks are at the end. Dates, titles and the full history are on LinkedIn.

§ 01

AI adoption at work

Which AI tool fits a piece of work depends on how its harness, environment and context work; when a task works in GitHub Copilot CLI, I turn it into a skill colleagues can run.

More on this topic

AI adoption · US health system · 2025–

Setting up colleagues to explore where AI becomes the best tool

I set up more than twenty colleagues with VS Code, GitHub Copilot, Python and Node.js, and showed them which Copilot fits which work. Microsoft 365 Copilot is for finding and using the organization’s own content, within each person’s own access; GitHub Copilot is for builds where I iterate on the same files many times. When a task works well in GitHub Copilot CLI, I transcribe it into a durable skill in the Agent Skills format (agentskills.io) so that colleagues can run it too; we have used these for PowerPoint presentations, HR headcount analysis and governance, risk and compliance (GRC) work.

AI at work · US health system · 2025–

Department newsletters, built with GitHub Copilot CLI

There was a need for internal newsletters, monthly and half-yearly, one for each department. I build them with GitHub Copilot CLI, and they come out engaging and professional, close to what a designer would produce. Microsoft 365 Copilot does not work for this, because its harness is not built for repetitive work where I keep iterating on the same files and folders under Git version control. GitHub Copilot in VS Code, or GitHub Copilot CLI, is the better tool because of how its harness, environment and context work. Without AI, we would not have thought of the task this way, and we would not have done it.

§ 02

Cybersecurity communication

My view, which I have not put into practice in my organization, is that awareness programs should use a careful, thoughtfully designed skills survey and request training completion by role, skill level and necessity.

More on this topic

Cybersecurity education · US health system · 2025

Cybersecurity comic strips for busy caregivers

In early 2025, there was a need for comic strips on cybersecurity topics, so that busy caregivers could understand a topic by glancing at a strip for a minute or two. The image-generation tools we had were not good at placing text in images, so I created the comic panels image by image with Microsoft 365 Copilot’s built-in image generation, then learned Figma to place the dialogue accurately on each panel. Some of the comic strips were used in a couple of newsletters. The rest were appreciated, but they never got a chance to be used at clinical sites, because of some limitations there.

Video · US health system · 2025

Product videos for an internal cybersecurity platform

I scripted and produced product videos for an internal cybersecurity platform, from storyboard to motion graphics in Premiere Pro and After Effects.

§ 03

Marketing technical products

When marketing cybersecurity products, we need to understand the pain points very clearly, and the solution we propose should clearly show its technical value.

More on this topic

Developer community · LoginRadius · 2021–2024

An open-source guest-author program

I ran the engineering blog as an open-source GitHub project with a paid guest-author program. Every post went through engineering review, and 144 merged pull requests on the repository trace the editorial work.

Ghostwriting · LoginRadius

Ghostwritten eBooks on data privacy and passwordless login

I turned in-house expertise on data privacy and passwordless authentication into long-form eBooks for customer education. They were published under an executive’s name, so they are not linked here.

§ 04

Research & notebooks

I learned how language models work by building and probing them. This section has a one-off red-teaming exercise from 2025 and notebooks from implementing and fine-tuning models.

Research · Red-Teaming Challenge: OpenAI gpt-oss-20b, Kaggle ·

Narrative and Reference Prompting as Safety Bypass Techniques

Red-teaming OpenAI’s gpt-oss-20b

My writeup for Kaggle’s Red-Teaming Challenge on OpenAI’s open-weight gpt-oss-20b. I tested whether the model’s safety layers could withstand indirect or nuanced prompting, using conversation-based strategies that mirror how real users might bypass safeguards, intentionally or not. With narrative-style dialogues and reference-based prompting, the model was coaxed into producing disallowed information.

Model
OpenAI gpt-oss-20b (open weights)
Techniques
Narrative prompting, reference prompting
Submitted
Aug 24, 2025

What it showed me

Naturalistic, conversation-based scenarios got further than confrontational prompts: putting a request inside a story, or approaching it through an indirect reference before getting more direct, made refusals less likely to trigger.

This was a one-off in August 2025. I have not done red-teaming since, the field has moved a lot, and I am not current on it. The exercise stays here for what it showed me about how a model responds to prompting, not as a claim to current red-teaming work.

This summary stays at the level of method. The full writeup, as submitted to the challenge, is on Kaggle.

Naturalistic, conversation-based scenarios made refusals less likely to trigger than confrontational prompts.