§ 01 — Operations Dossier
Six standing research divisions. Active project files. Some public, some not. Inclusion does not imply completion, sanity, or institutional support.
§ 01.A — Research Divisions
What actually caused what? Statistical models find correlation; causal inference asks the harder question. The group works through Pearl's hierarchy, runs do-calculus exercises, and applies counterfactual methods to real datasets — currently a clinical decision-making collaboration with Mass General.
Model interpretability is no longer optional. The group surveys and implements XAI methods — SHAP, LIME, integrated gradients, attention rollouts — and audits real models. A 2019 project from this group was deployed at Squark.ai. Multiple papers in preparation.
Where information design meets machine learning. Builds interactive visualization tools that surface what models are doing — and lets domain experts disagree with them in a structured way. Cross-listed with the Information Design and Visualization program.
Production models drift, fail silently, and lie about their confidence. The group builds instrumentation: drift detectors, calibration audits, fairness monitors. The unsexy work that decides whether a model survives contact with reality.
Competition-driven learning. Pick a Kaggle competition each semester, work it as a team across feature engineering, modeling, and ensembling. Currently working the RSNA Pulmonary Embolism Detection challenge.
Generative art research. Named for the painter who left the Paris market to make work the market wasn't asking for. Studies aesthetic judgment as a research problem — what makes generative output good rather than merely novel. Outputs: exhibitions, papers, the occasional album.
§ 01.B — Active Project Files
A collaboration between designers and developers — bridging the gap between organically-generated design concepts and what models can actually produce. Built by a team that took RISE finalist standing in 2019. Still iterating. Contributors: Tanvi Modi, Yuhang Li, others.
RSNA Kaggle challenge work. 3D CT volumes, class imbalance, competition-grade ensembling. Cross-pollination with the Causal Inference group's clinical work at Mass General.
Open-source XAI methods bundle maintained by the Explainable AI group. SHAP, LIME, integrated gradients, attention rollouts, and audit tooling — all in one place.
Real-time drift detection prototype from the ML Observability group. Planned for open-sourcing once it stops embarrassing the team. Currently deployed against three internal partner datasets for validation.
Creative-coding tool built by the Gauguin group. Used internally for exhibitions and the occasional album cover. Diffusion + aesthetic evaluation loops.
██████████████████████ in █████████ settings. Details withheld pending review.
§ 01.C — Intake
Skunkworks has always grown by people showing up with something specific to investigate. Submit a two-page proposal year-round. We respond within fourteen days.