Private by design
Compare permitted submission metadata with your research profile using an embedding model that runs entirely in this browser.
Step 1
Each item is embedded separately so matches can point back to its strongest influence.
Step 2
CSV and JSON stay in memory unless you explicitly enable local persistence.
Expected fields: paper ID, title, and ideally abstract, topics, track, preference.
Step 3
The model will use WebGPU when available, with a WASM fallback.
Step 4
Read the paper metadata and make your own judgment before recording a preference.
Interested papers boost similar candidates; not-interested papers apply a stronger similarity penalty. Conference topics are not used for scoring.
No rankings yet.
Local persistence
Enabled by default. Profile text, paper metadata, scores, statuses, and notes are stored only in this browser’s IndexedDB. Turn it off to remove the saved copy.
This session will be saved locally.
Export
Use blank or an integer from −20 to 20. Conflict-marked papers are omitted; declare conflicts through the official conference process.
Technical overview
The ranking is a local semantic comparison, not an assessment of reviewer fitness.
Your weighted expertise items and each paper’s title and abstract are normalized and split into overlapping chunks when they exceed the model’s practical context size. Imported conference topics remain available only as display/filter metadata.
The app uses the quantized all-MiniLM-L6-v2 sentence-transformer model. It produces 384-dimensional vectors in a Web Worker, using WebGPU when available and a local WebAssembly fallback otherwise. Model files are served with the app.
Chunk vectors are mean-pooled and normalized. Cosine similarity compares each paper with every profile item. The final score blends the weighted profile centroid (70%) with the strongest individual expertise match (30%).
The strongest matching profile items explain each base score. Interested decisions add a small similarity boost, while not-interested papers and manual negative phrases apply stronger penalties. Similarity never becomes a bid automatically, and conflicts remain part of the official conference process.