Show the reasoning
A prediction nobody can inspect is a prediction nobody can challenge, so explanation is built into the model.
Grad-CAM revealed where the 40X breast cancer model was looking when it got a malignant slide wrong.
About
AI/ML engineer building systems that people can trust.
I'm an AI/ML engineer who builds systems people can trust: models that explain their decisions, pipelines that fail safely, and analysis that is honest about what the data can and can't say.
At DataCompass, I build AI for a clinical trial platform, and two of my systems work as a pair. The first is a profile automation that enriches the platform's knowledge base, producing 85 structured company profiles per cycle and replacing a manual research process that took 42 hours. The second is an AI agent that puts that knowledge to work: contract research organizations (CROs) use it to fill out RFPs, drawing answers from the enriched knowledge base. I also manage the Supabase data infrastructure underneath both, including the pipelines and quality checks that keep the data the agent relies on accurate.
My graduate work at the University of Denver focused on deep learning, statistical modeling, and LLM systems, applied to medical imaging, climate forecasting, and public-interest analytics. I came to AI from behavioral neuroscience, so I treat interpretability, model risk, and environmental cost as design requirements rather than afterthoughts.
How I approach responsible AI
It only counts if it shows up in the work. Each principle points to a project where it changed a decision.
Report what the data can't say
Visually suggestive isn't statistically significant. Effect sizes and limits go next to the headline.
Fashion transparency explained only 14% of sustainability performance, reported as a weak link.
Weight the errors that matter
Not every mistake costs the same. Metrics should reflect who gets hurt when the model is wrong.
Malignant recall, not overall accuracy, was the target metric for the histopathology model.
Count the cost
Compute has an environmental and financial bill. Use the smallest model and simplest stack that does the job.
Daily Companion makes at most one small Claude Haiku call per day, and runs everything else locally.
Have a problem worth modeling carefully?
I'm glad to talk about interpretable ML, LLM systems, or a dataset that deserves better.