Daylor Williams

AI Product Builder + ML Researcher

Building and studying AI systems from model behavior to real-world deployment.

Selected Work

01 — Four systems

01

Agentic Prior-Art Platform

UCLA Technology Development Group · AI Product Intern · 2026

Every invention disclosure needs a prior-art search across global patent literature — slow, manual work that gates licensing decisions.

Owned

  • Customer discovery with licensing officers
  • Product requirements for the agent system
  • Golden evaluation datasets for grading agent output
  • Agent failure analysis and taxonomy

Result

Co-building an enterprise agentic system for invention evaluation and prior-art research, evaluated against golden datasets built from real disclosures.

  • Python
  • APIs
  • Azure
  • EPO infrastructure
  • Agent evals

02

DINOv3 Representation Research

Computational Vision & Learning Lab, UCLA · First author · targeting ICLR 2027

Vision models encode what objects are and how they relate in the same embedding — entangled, and hard to inspect or control.

Owned

  • First-author study: design, experiments, writing
  • Embedding-space analysis of DINOv3 features
  • Causal intervention and representation steering

Result

Decomposed DINOv3 representations into object-identity and spatial-relationship components, with causal interventions that steer behavior along each axis independently.

  • PCA
  • LDA
  • Linear projections
  • Representation steering

Full write-up →

03

AI Workpaper Conversion

PwC · Workflow design · 2025

Converting audit workpapers between formats consumes hundreds of billable hours per engagement, all of it manual.

Owned

  • End-to-end AI-enabled conversion workflow design
  • Accuracy checks and full traceability
  • Human-review controls at each step

Result

Modeled ≈480 hours and ≈$150K in savings per engagement; earned engagement-partner sponsorship for rollout planning.

  • Workflow design
  • QA controls
  • Human-in-the-loop review

04

GRAPES Search+

Independent → UCLA Graduate Division · Founder · ongoing

UCLA's graduate funding database decays silently — listings expire without notice, and students search dead records.

Owned

  • Conceived the product independently
  • Audited all 625 funding records by hand
  • Built the validation pipeline

Result

Found only ~350 of 625 records still active — a data-quality problem nobody had measured. Now working toward implementation with Graduate Division leadership.

  • Data auditing
  • Validation pipeline
  • Product discovery

About

02 — A short introduction

I'm a cognitive science and computing student at UCLA who builds AI products and studies what's happening inside the models they depend on.

The two halves are one job: you can't ship an agentic system for consequential decisions — patent evaluations, audit workpapers, funding searches — without knowing how it fails. So my product work runs on golden datasets and failure analysis, and my research asks what vision models actually encode.

I'm interested in the gap between what AI demos and what AI can be trusted to do.