Daylor Williams

Owned

• Solo build: lab design, experiments, site, in 24 hours
• Autonomous lab of AI agents: PI, pods, analyst, skeptic, reviewer
• Pre-registered hypotheses and a synthetic test environment
• Statistical analysis with confidence intervals on every claim
• React/TypeScript results site, deployed on Vercel

Result

Ran four research loops on Claude Haiku and Qwen agents studying when they resist correction. Most hypotheses came back null, and the lab said so: 0/90 agents persisted with an invalidated plan. Its first real effect: an advisory majority saying "skip it" cut verification by 30 points. Result-to-decision time fell from 60 minutes in the first loop to 6–11 minutes in later loops.

Owned

• Trace schema and comparison model
• Deterministic alignment (weighted Needleman-Wunsch)
• Field-level behavioral diffs
• React/TypeScript trajectory interface
• Claude Code transcript adapter

Result

Aligns two runs, collapses identical behavior, and shows the first observable difference in tool calls, arguments, results, errors, retries, and stopping. Validated on real Claude Code sessions without requiring manual comparison of raw transcripts.

Owned

• First-author study: design, experiments, writing
• Embedding-space analysis of DINOv3 features
• Causal-intervention experiment design (in progress)

Result

Removing the object subspace from a frozen embedding leaves a residual that still carries spatial relation: 93.3% cross-object generalization against 53.3% from raw embeddings, and the signal transfers to 400 photographed real-world objects at 97.9%. Currently testing whether interventions on the relational component produce the predicted changes in spatial judgments.

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 agentic prior-art system, graded against golden datasets built from real disclosures across sparse, crowded, and inconclusive technology fields. The cases it gets wrong help define what should remain with a human reviewer.

Owned

• Sole author: design, build, store release
• Manifest V3 extension that renders inside Google Calendar
• Trip detection from all-day events and their locations
• NOAA forecast blend for US locations, with caching
• Privacy-first: no servers, accounts, or analytics

Result

Each day's forecast appears under its date in Calendar and switches to the destination city when you travel. US forecasts use NOAA's blend because the default model ran about 10°F hot for coastal LA. It's built without innerHTML to satisfy Calendar's Trusted Types policy, and it reads no calendar data through Google's API.

Owned

• Designed the conversion workflow and its review steps
• Accuracy checks against the source workpaper
• Human-review controls at each step

Result

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

Owned

• Sole author: product, design, data pipeline
• September 2026 redesign of the search experience
• Audit of every listing against its source
• Python pipeline with scheduled GitHub Actions updates
• Verified / unverified labels on every listing
• Working with Graduate Division leadership on adoption

Result

Rebuilt as a search-first funding finder: plain-language search with related terms, filters by deadline, award type, field and season with live counts, saved awards, and shareable search links. Passed deadlines roll forward to the expected next cycle instead of disappearing. The redesign audit removed 22 dead listings, updated 116 deadlines, replaced 25 broken links and added 88 missing awards, bringing it to 400+ live opportunities. Now working with Graduate Division leadership on bringing it into official use.