Open-source infrastructure for inspectable research.

An open-source Python package for detecting explicit boycott discourse in brand-level text data. The classifier looks for boycott language, purchase-avoidance calls, brand-specific boycott hashtags, or campaign terms near the target brand, and rejects general negativity that is not explicitly about boycotting.

Built from the construct-measurement pipeline behind the social-media boycotts and market-attention paper.

An open-source Python toolkit for analysing promotional disclosure style and explicit superiority rhetoric in earnings calls, annual reports, speeches, startup pitches, shareholder letters, and other strategic communication.

The tool separates broad promotional style from stricter narcissistic-distinctiveness claims and preserves sentence-level evidence behind every score.

An open-source Python pipeline for evaluating temporal transportability in suicidal-ideation prediction. The toolkit reconstructs the 176,957-respondent employed-adult NSDUH analytic sample, tests models across all 2015–2023 train-year and test-year combinations, and generates paper-ready validation artifacts.

Built from the machine-learning suicide-prediction paper. It is a replication and governance research tool with an independent audit protocol; it is not a clinical diagnostic system, individual risk scorer, or employer screening product.

An open-source model-release repository containing a fitted suicidal-ideation prediction model, feature schema, CSV scorer, Python API, synthetic examples, model card, validation checklist, governance guidance, and a reproduced fresh-data validation workflow on NSDUH 2024.

Built from the machine-learning suicide-prediction paper and separated from the paper reproduction repo. It supports local validation and governed methodological prototyping; it is not a clinical diagnostic system or an automated employment decision tool. The 2024 validation reproduced on 20,588 employed respondents with useful temporal transportability (AUC 0.830), but the packaged threshold requires local recalibration and one predictor was unavailable in the public-use file.