Taylor Clements, PhD · AI engineer and data scientist
The traditional way of building software, one layer of the stack at a time, is fading. What's left is judgment: knowing which questions to ask, and how to work with AI to answer them well. I build AI systems that hold up under real use, and I measure them before I trust them.
Selected work
An agent that answers routine calls for a home-services company, books appointments in a real scheduling database and hands everything else to a person. Before building it, I classified all 420 sample calls to learn why people call. Code, not the model, decides identity, bookings and handoffs.
An assistant inside an open-source medical record that gives an overnight resident the patient context they need, sourced from the actual chart, uploaded lab reports and clinical guidelines. Every claim cites its source, and an eval set blocks every code change that fails it.
Four agents that red-team the clinical co-pilot from the outside, the way a real attacker would. An independent judge scores every attack, and nothing is published without a human approval that no model can set. It found real problems in my own co-pilot, and it confirmed the fix.
More projects
A coach dashboard that builds safe, personalized workouts. A language model can propose an exercise, but only a knowledge-graph check can allow it.
Code checks and a judge model score each AI-written progress note, repair what they can, and pass, flag or block it for a human coach.
A platform for wiring AI agents into visual workflows, with tools, memory, approvals, cost caps, live monitoring and a Slack connection.
Approach
Before building the call center agent, I classified every sample call to learn why people call. That showed even a perfect agent could handle only 61% of them, so I aimed at what mattered instead of a number it could never reach.
Every project ships with its evals: golden sets that block a merge, held-out test sets and red-team suites. The numbers on this site come from committed result files you can check.
Models propose; code decides. Identity checks, bookings, safety rules and human approvals live in plain code that no prompt can talk its way around.
I track spend on every run and design for it: local models where they're good enough, paid models only where they earn it, and hard budget caps on anything public.
About
I spent nearly ten years teaching math and leading schools, then earned a PhD in Education Sciences from the University of Kentucky. Teaching taught me the skill that matters most with AI: asking a clear question and checking the answer.
Today I'm an enterprise analytics translator at Sysco, turning business problems into models and decisions, and I build AI coaching and retrieval pipelines for tutors at Cignition. I trained in AI engineering at Gauntlet AI and I'm finishing an MS in Data Science at Eastern University.
I still work with schools and school leaders through Learning V. School.