Brayton Hall
Data Scientist · Language Models · Philosophy and Literature
For the past several years I have worked on language models at large banks, mostly on the problem of usefully leveraging those models in a safe and scalable way. Models from Anthropic and OpenAI are becoming superhuman, but I still believe there's an art at play that goes beyond prompt engineering.
The deep math, the mechanisms of transformers and reinforcement learning, and the recursive loop between human behavior and AI acceleration are a critical new region of interest somewhere between math and philosophy and computer science, which is where I've lived for years. I left Capital One and moved back to NYC in August 2026: check back soon for some exciting updates...
Before this
I studied philosophy, metalogic and number theory at UNC Chapel Hill, before teaching logic and statistics at Virginia Tech. My research was on the history of mathematical logic and the philosophy of language, particularly in the early 20th century, with a focus on the philosophy of Ludwig Wittgenstein. The subject of my thesis was on the scientific explanation of events in history, and ranged over canonical twentieth century texts including Wittgenstein's Philosophical Investigations and Thomas Kuhn's The Structure of Scientific Revolutions.
Off the clock: long runs (the occasional marathon), New York by ferry, movies at the Metrograph, the piano, writing, and slow, happy progress through Joyce and Proust.
Last touched by hand on .
Work & Education
Soon.
Lead developer of the company’s second production generative-AI use case: summarizing 1M+ highly sensitive documents for executive committees (≈$20M/yr). Model development, compliance, and monitoring for the fraud model covering every Capital One card (≈$400M/yr in prevented loss).
Built and monitored the bank’s primary abstractive-summarization pipeline for calls between agents and customers, trained on 600K agent notes and 150M transcribed conversations (≈$10M/yr), with custom summary-quality metrics.
ML on infant-brain fMRI (100K+ features) for early-intervention prediction: interpretable models, Shapley values, feature extraction.
Metalogic, number theory, philosophy of science and language. Taught logic and statistical inference.
Projects
Obviously old fashioned now in the transformer age, but I was proud of them in 2020.
Semantic Search
Search classic novels by connotation and misremembered quotes rather than exact fragments. Doc2Vec embeddings over Project Gutenberg.
Mapping Ulysses
Vectorizing each chapter of Joyce and plotting the novel’s drift through its own semantic space. Data science as literary criticism.
Summarizing War and Peace
Few-shot fine-tuning of pre-trained transformers for long-document summarization.
codeTwitter Topic Modeling
LDA over 100K pandemic-era tweets: finding the distinct communities and their sentiment.
code