Dr. Daniel Wooten
Principal Data Scientist | Card AI Research & Development
Daniel Wooten
About Me
I take research models into production and then stay for the part where they have to keep working.
Seven years of building, shipping and then living with AI/ML models have left me most useful in the gap between research and production — the place where a method that works beautifully in a notebook still has to survive a review board. My current work at Capital One is generative AI and transformer architectures pointed at risk and operations, problems where a single percentage point is worth eight figures.
I've taken projects the whole way, from a whiteboard sketch to a model a federal examiner can pull apart line by line, and it's that last stretch I find most interesting — the point at which an algorithm stops being a clever observation about patterns in data and becomes something a business actually decides with, auditing, defending and all. The clever part is rarely the hard part.
Current Impact
Principal Data Scientist at Capital One
(Card AI R&D)
Project OWL: LLM on Call Transcripts
I built an LLM system to grade call-transcript analyses that people had been grading by hand. Iterative prompt engineering and meta-prompting brought it to human-level agreement. That is worth $30 million a year to the bank.
I built an LLM system to grade call-transcript analyses that people had been grading by hand. Iterative prompt engineering and meta-prompting brought it to human-level agreement. That is worth $30 million a year to the bank.
Project Genesis: Acquisitions Risk
I lead our work putting transformer architectures on raw credit bureau data to predict charge-off risk. Read as a sequence rather than a pile of aggregate features, payment history gave the proof of concept enough signal to beat the gradient-boosted models that are the standard here (on the order of $15 million a year, if it holds at scale).
I lead our work putting transformer architectures on raw credit bureau data to predict charge-off risk. Read as a sequence rather than a pile of aggregate features, payment history gave the proof of concept enough signal to beat the gradient-boosted models that are the standard here (on the order of $15 million a year, if it holds at scale).
From the heart of AI
Ph.D. in Nuclear Engineering, UC Berkeley
The large-data habit started at Berkeley, where my thesis ran on datasets north of a terabyte and I was an original member of the Berkeley Institute for Data Science, back when machine learning and computational engineering were still being introduced to each other. In concrete terms that thesis is a 30,000-line extension to a Monte Carlo reactor physics code written in C. That's the nuclear side of my background, and it has a page of its own.
Foundational Experience
AI/ML Engineer at Perceptronics Solutions
Bayesian inference systems there, and real-time tracking across a whole city. Much of the job was sitting with subject-matter experts and turning what they knew but could not quite write down into something an algorithm could use. Multiple-Hypothesis Tracking and Generalized Multi-Bernoulli filters, mostly, for environments that offer you no training set to learn from.
What do I bring to your team?
Three things, if I had to pick.
Value You Can Count
I like models whose worth can be written down as a number. The two above come to eight figures a year between them.
Making the Case Upstairs
I've argued for transformers and LLMs in front of executive leadership. That's a different skill from building them and, I'd argue, the harder one.
Paper to Production
Berkeley trained, and still comfortable taking an architecture straight off a preprint and standing it up somewhere it has to stay standing.