Abstract
Two systems sit at the center of my thesis, and both spend their working lives deciding what to trust. My technical project is a radar-based object detector for autonomous racing that must separate real vehicles from noise, and my STS research examines the “accredited investor” standard, the wealth threshold that decides who may legally invest in U.S. private markets. The projects share no subject matter, but they share a structure: each is a gate that judges a hidden quantity through an imperfect proxy, and each is defined less by the cases it gets right than by the errors it makes. A detector that hallucinates a car from clutter and a rule that admits an unqualified investor from a bank balance are the same kind of failure. Studying both together taught me that the design of any such gate is never purely technical, because choosing what counts as an error, and who pays for it, is a decision about values.
The technical portion of my thesis adapted a self-supervised learning method to automotive radar. I built on AD-L-JEPA, an architecture designed for dense LiDAR point clouds and ported its core idea to the far sparser and noisier radar domain: mask out regions of a bird’s-eye-view map and force the model to predict the hidden features from the surrounding visible context. Before any labeled training, this pretext task pretrains the detector’s sparse three-dimensional backbone on roughly 36,000 unlabeled radar frames; the detector itself is a radar-only CenterPoint pipeline. In a controlled comparison that held the dataset, configuration, and training schedule fixed, my strongest pretrained model raised precision from 0.73 to 0.77, lifted F1 from 0.78 to 0.81, cut false positives per frame from 0.26 to 0.21, and improved mean localization error from 0.34 to 0.32 meters, all while recall stayed essentially flat. It did not improve every metric, so I claim only what the evidence supports: the radar pretext task taught the backbone a cleaner sense of radar-supported versus empty space, which made the detector less likely to invent objects from clutter.
In my STS research, I asked why the accredited investor standard has survived more than thirty years of documented criticism. Nearly every economist agrees that wealth (a million-dollar net worth or a $200,000 income) is a poor proxy for the financial sophistication the rule claims to measure, yet the threshold endures. Rather than treat this as a measurement error awaiting better data, I treated the rulemaking process as a contested policy subsystem, combining Sabatier’s Advocacy Coalition Framework, Entman’s framing theory, and Stigler’s theory of regulatory capture to code primary-source comment letters and the legislative record of the 2025 INVEST Act. My analysis revealed that the two coalitions publicly fighting over the standard quietly agree to preserve wealth-based gating, differing only over its calibration; an examination-based alternative fails not on its merits but because it is the one reform that would strip verification authority from every incumbent positioned to block it. The proxy persists as a political equilibrium, and reform must treat the assignment of that authority as a first-order design decision.
Placed side by side, the two projects describe the same problem from opposite ends. My technical work let me tune a gate mathematically, trading false positives against recall as if accuracy were a neutral target. My STS work showed why, in a human system, that target is never neutral: the errors a wealth test produces are not bugs to be optimized away but benefits to the actors the rule protects. Seeing the technical, organizational, and cultural layers at once is what made this visible, and a classifier, whether it screens radar returns or investors, always encodes a decision about whom to protect and whom to expose to risk. That is where I locate ethical responsibility in engineering. The most rigorous detector and the most defensible regulation are still, underneath, choices about how to distribute the cost of being wrong, and an engineer who treats those choices as merely technical has already made them without noticing.
