Universe

Lyra Motion Intelligence Lab

Hi, I'm Lyra.

Stylised 3D floating head of Lyra

My current interests span data science, multimodal sensing, and spatiotemporal learning, with applications in human movement and event understanding.

Research focus

Reliable Multimodal
Spatiotemporal Intelligence

My research focuses on multimodal sensing and spatiotemporal learning for human movement analysis and event understanding. A core emphasis is placed on ensuring consistent, robust performance—guaranteeing reliability across diverse individuals, recording sessions, hardware devices, and operating environments.

Wearable & Mobile SensingComputer VisionAudio IntelligenceWireless SensingSensor FusionEvent & State Understanding

Research projects

Research and technical work.

My work examines how reliable representations of human movement and events can be recovered from heterogeneous, imperfectly observed data.

I develop and evaluate multimodal sensing systems across inertial wearables, UWB and wireless sensing, computer vision, identity-aware tracking and trajectory modelling, with an emphasis on robust validation and end-to-end experimental design.

Three-radar UWB fall localization and shortcut-audit pipeline

Wireless sensing & temporal localization

Shortcut-aware UWB Fall Localization

Reproduced an ABC 2027 UWB baseline and built a three-radar localization pipeline with subject-disjoint evaluation. Shortcut controls showed that much of the tree-model gain survived temporal motion destruction, exposing a robustness–accuracy trade-off rather than claiming clinical validity.

UWB radarTemporal localization
Fold-specific change in Kendall ranking agreement for SynthShift

Wearable & inertial sensing

SynthShift

Built a subject-disjoint evaluation framework for cross-device IMU domain shift. Across 60 test cells, calibration reduced failure-prediction MAE from 0.109 to 0.067 and achieved mean model-ranking agreement of τ-b 0.677.

IMUDomain shift
Video
Detect
Track
Fuse identity
Validate

Computer vision & identity tracking

TrackID3x3

Built an end-to-end vision pipeline covering player detection, court projection, tracking, team classification, jersey evidence and global identity assignment. Grouped video validation and TI-HOTA keep model selection separate from frozen challenge inference.

DetectionReID
Strong-baseline validation for the Defensive Access metric

Spatiotemporal tracking

Defensive Access

Built a pre-pass choice-set model to measure how defender geometry changes passing access. Match-held-out tests found a small log-loss gain but no ranking improvement, and timing falsification limited the strength of the claim.

Tracking dataCounterfactual geometry
Revision evidence summary for temporal basketball defence modelling

Spatiotemporal event modelling

Beyond the Snapshot

Audited 1,211 shots to test whether temporal defensive context improves shot-quality models. The added context produced no reliable predictive gain, while revealing defender-identity changes in 36.0% of one-second windows.

Event alignmentUncertainty audit
Music Ring virtual instrument and gesture console

Wearable sensing & interaction

Music Ring

Built a BLE and IMU gesture platform that turns smart-ring motion into musical control. The system integrates a FastAPI session, web and phone interfaces, local audio playback and an optional spatial display.

BLEIMU gestures

Selected applied work

Sports decision intelligence

Project / 01Forecasting

KPL · Esports

OpenKPL Analytics

A reproducible tournament intelligence system combining audited historical series, leakage-safe team-strength modelling, Elo benchmarks, and Monte Carlo progression forecasts.

PythonEloMonte CarloNext.js
Project / 02Forecasting

NBA · 2026 Draft

NBA Draft Intelligence System

A probabilistic draft board combining multi-source mock consensus, betting markets, team-fit modelling, calibrated uncertainty, and Monte Carlo draft simulation.

PythonMonte CarloData PipelineNext.js
Project / 03Forecasting

FIFA · 2026 World Cup

World Cup Forecasting Engine

A probabilistic football platform trained on 49,000+ international matches, combining Elo, Poisson and Dixon–Coles scoring models with full-tournament simulation.

EloDixon–ColesPoissonSimulation
Project / 04Research

NBA · Spatial shot analysis

Basketball Range Intelligence

An interactive shot-location research tool for comparing how NBA players' range and efficiency evolve across seasons, with chronological replay, density fields, and distance-based shooting profiles.

Spatial AnalyticsCanvasEChartsPython
Project / 05System

Football · Live event intelligence

Football Director OS

A natural-language match monitor that compiles viewer questions into validated temporal rules, detects patterns in a live synthetic event stream, and returns multilingual, evidence-backed explanations.

Temporal EventsExplainable AIFastAPIReact

Research · build · learn in public

Interested in the
same questions?

Let's collaborate