Things I’ve built or studied
Doctoral work on active noise control, quantitative research on Indian equities, and a few things that fit neither. Each page states what the question was, how it was measured, and where the answer stops being reliable.
- Meta Quant — a machine that searches for trading strategies A constrained language for strategies, an automated search over it, and an honest account of what has not worked yet →
- Making chart patterns measurable Turning an informal visual vocabulary into detectors, and detectors into language a model can use →
- Pattern reuse, from file systems to neural networks Testing a 2018 compressed-filesystem idea against raw binary, and what bit alignment revealed →
- Quant investigations into Indian equities (NSE/BSE) Custom UIs, plots, and statistical studies of Indian equities, started Jul 2024 →
- A toolbox for collecting human preference data Instrumentation for comparing preference elicitation methods →
- Measuring whether image generators can depict an emotion A within-subjects study of emotional alignment, and a rescue archive of deprecated models →
- Characterising ANC headphones through the playback signal Reframing ANC evaluation around the playback signal, not the noise →
- Eighty listeners, 360 degraded sentences each, and a mostly negative result A CRM intelligibility and NASA-TLX effort experiment on shallow-envelope speech chimeras →
- Ten instrumental metrics, two corpora, and a jury that cannot agree FES, GPSMq, HASPI, HASQI, PESQ, PEMO-Q, STOI, ViSQOL, wSTMI and SRMR applied to ANC playback signals →
- A simulated auditory nerve as a virtual listener for ANC headphones Driving the Zilany/UR EAR auditory cascade with real headphone recordings to get a perceptual fingerprint without recruiting people →
- Building a modulation transfer function for an ANC headphone Eight envelope statistics, swept across modulation frequency, on four real devices →
- Designing and teaching a graduate course from scratch Eleven lectures of machine learning and game theory, plus the evaluation data →
- Identifying users by how they type — a functional-link neural network for keystroke dynamics A from-scratch FLANN classifier for typing-rhythm biometrics, benchmarked against MLP, random forest, k-NN and trees →