ML CompetitionPortfolio
Selected Medal Cases, Retrieval Work, and Public Artifacts
A selected archive of competition work rather than a generic Kaggle dump: one silver-medal case, two bronze-medal cases, and a larger retrieval-and-embeddings project surfaced as a separate technical artifact.
A Small, Real Competition Record
Primary highlight built around the exoplanet signal-processing write-up.
One public RSNA case and one private unpublished case represented honestly.
The EEDI retrieval project broadens the archive beyond medal entries without muddying the record.
Different Cases, Different Kinds of Evidence
Exoplanet Signal Extraction
The primary silver-medal case centers on a calibration-first pipeline for noisy sensor cubes, where smoothing and compact signal engineering mattered more than a larger model.
Lumbar Spine / RSNA
The public bronze-medal case is framed around task decomposition: splitting heterogeneous spine targets into more coherent specialists instead of forcing one global classifier.
Private Bronze Case
A second bronze-medal solution is represented here honestly as unpublished/private work. It is included for portfolio truthfulness, not inflated with fake public artifacts.
EEDI Retrieval + Embeddings
A larger competition-related project built around retrieval and embeddings. It belongs in the archive as a substantial technical artifact, but not as a medal entry.
What Repeated Even When the Cases Changed
Decompose heterogeneous targets
The strongest results often came from refusing to treat every label family as one unified problem when the structure clearly suggested specialization.
Simplify the signal first
In scientific or noisy-data competitions, calibration, denoising, and representation choices often matter more than model size alone.
Use ensembles pragmatically
Competition work benefited from ensemble thinking, but only when it reflected distinct useful views of the problem rather than stacking models for its own sake.
Treat retrieval as a real system
The larger EEDI work belongs here because it extends the portfolio from tabular and imaging tactics into retrieval-and-embeddings design.
Public Things a Reader Can Actually Open
The wider competition profile and public notebook/archive surface.
OpenSilver-medal write-up focused on calibration, signal extraction, and smoothing.
OpenThe competition lab note that captures the denoising and smoothing lesson behind the silver result.
OpenPublic repository for the lumbar-spine bronze-medal case.
OpenThe lab note that frames the bronze-medal result as a decomposition-first classification strategy.
OpenLarger embeddings/retrieval competition artifact, separate from the medal record.
Open