Conference affiliation: Senior Data Scientist · Lease End · 2026
On this page
Dan Bjornn is a cognitive neuroscientist turned data scientist who builds customer-facing AI systems and studies the operational costs of deploying them. As a senior data scientist at Lease End, he helped develop an automotive-finance sales agent that generated a reported $12 million in revenue at 50 times its investment—and subsequently rebuilt the system when fine-tuning became an obstacle to improvement.
From behavioral research to production AI
Bjornn studied psychology at Brigham Young University, completed his doctorate in 2018, and specialized in cognitive and behavioral neuroscience. His research encompassed memory, decision-making, user experience, and security behavior; at BYU’s Neurosecurity Lab, he investigated how people assess digital risks.
He later joined NeuroID as an early employee, helping establish the technical foundations of a behavioral-analytics company focused on distinguishing legitimate users from potential fraud. Experian acquired NeuroID in 2024. At Lease End, Bjornn applied that behavioral perspective to an SMS agent connecting drivers approaching the end of their auto leases with financing options and sales representatives.
The hidden cost of a successful AI agent
The messaging application, developed in late 2024, initially used a vector database and retrieval-based customer-intent classification. Bjornn subsequently built a supervised fine-tuning pipeline incorporating LLM-generated labels, manual review, holdout evaluations, and synthetic examples.
Despite its commercial success, the system repeatedly misread conversational nuances: confirming a future appointment or offering a routine greeting could trigger an unwanted immediate call. Fixes required collecting and validating examples, retraining, and checking for regressions—a cycle lasting approximately a week.
The calcification tax: Fine-tuned models can harden an application around specific providers, model versions, training formats, and workflows, making upgrades increasingly expensive.
Model-agnostic agent architecture: Replacing the classifier with reusable skills, tools, curated evaluations, and stronger contextual instructions allowed updates to be deployed as Markdown files through Amazon S3, reducing fixes from roughly a week to under an hour.
Total operating cost over token cost: Although stronger models increased per-message API spending, faster iteration, reduced maintenance, and fewer customer-facing failures lowered overall operating costs.
Bjornn treats fine-tuning as a last resort, appropriate primarily when privacy, offline operation, or comparable constraints prevent capable frontier models from meeting the requirement.
