Sonny Fullerton

Machine learning, signal processing, and statistical modelling for difficult problems.

Currently exploring AI-generated music detection, prediction-market derivatives, and representation learning.

EmailGitHubLinkedIn

Current questions

  • Can AI-music detectors generalise to generators they have never seen?
  • How should probabilities diffuse when the state space is bounded?
  • Which detection claims survive codec laundering and distribution shift?
  • Can false-positive guarantees hold across genres and distribution shift?

Research

Detector

Cross-generator forensic detection of AI-generated music.

A research system for testing whether AI-music detection can generalise across unseen generators, ordinary re-encoding, and changing musical domains, with explicit control of false accusations against human work. The implementation remains private while commercial options are evaluated.

Mean clean AUC 0.993; mean recall 87% at a 1% human false-positive operating point.

Meridian

What changes when an option's underlying is a probability?

A focused study of European options on bounded probability assets, from the logit transform to the martingale correction it requires.

  • Research note
  • Notebook
  • Calculator
  • Source

Projects

Florescence

A real-time audio plugin built from the signal path up.

An audio-plugin project spanning DSP, low-latency real-time code, and interface design.

  • Plugin
  • Architecture
  • DSP write-up
  • Source

Sphera

Making a mix spatial, immediate, and legible in real time.

A real-time 3D audio analysis tool, supported by the Wharton Venture Lab Innovation Fund.

Writing

Does randomness fundamentally exist?

A metaphysical argument separating epistemic uncertainty from ontic and fundamental randomness, moving from Laplace and Leibniz through quantum measurement, Bell's theorem, and emergence.

Final paper for Introduction to Philosophy at the University of Pennsylvania.

  • Probability
  • Uncertainty
  • Causality
  • Emergence
Read paperDownload .docx

Research notes

Detector

  • Cross-generator evaluation
  • Codec and transformation robustness
  • Low-false-positive calibration
  • Song-disjoint benchmark design
  • Confounds and failed shortcuts

Meridian

  • Why Black–Scholes does not fit
  • Logit and sigmoid transforms
  • Itô drift correction
  • Naive-versus-corrected bias
  • Put–call parity checks

Notes will be published as they become useful to someone other than their author.

Other work

Homophily under model substitutionA partial replication of He et al. (2026) on Claude Haiku 4.5, with a live dashboard and research note.WagerlyA social prediction market with parimutuel settlement and a zero-sum ledger.SonnateAn AI assistant studio that deployed booking and engagement systems for SMB clients.