AI Skunkworks · Northeastern University · est. 2019
A student-led AI research and build group at Northeastern University.
Northeastern graduate students, researchers, and mentors who want to do real AI work beyond the coursework. Founded in 2019 under faculty advisor Prof. Nik Bear Brown, revived in January 2026, and affiliated with Humanitarians AI. Membership is open to anyone at Northeastern willing to do the work.
§ 01 — Premise
Field Manual · Entry 001
A skunkworks project is one built by a small, loosely-structured team whose only job is radical innovation. The term comes from Lockheed's secret aircraft division — they built the U-2 in eight months in 1955 in a rented circus tent next to a plastics factory.
Ours is the student version. No actual skunks involved. What translates is the operating model: small teams, short loops, real stakes, and the freedom to be wrong in interesting ways.
§ 02 — Operational Layers
LAYER · TOOLING
TIER 01 — NECESSARY, NOT SUFFICIENT
Frameworks, libraries, training pipelines, deployment. The table stakes. Required, but not the reason you joined a skunkworks. Machines are already superhuman at this. Operate the forklift; don't compete with it.
LAYER · RESEARCH
TIER 02 — WHERE WE LIVE
Six active research groups: causal inference, explainable AI, observability, visual analytics, generative art, and competition ML. Student-led, faculty-advised, mentor-supported. Real papers, real deployments, real failures.
LAYER · JUDGMENT
TIER 03 — THE IRREDUCIBLE PART
The thing AI cannot supply: problem formulation, taste, the call about what's worth your semester. The reason a skunkworks exists at all. Practical wisdom under real stakes — no model has any.
§ 03 — Active Divisions
Counterfactuals, do-calculus, and the difference between correlation and what actually caused what. Weekly reading group plus active project work — including a clinical decision-making collaboration with Mass General.
Model interpretability methods, audit tooling, and making black boxes admit what they're doing. Multiple ongoing papers. Deployment lineage to Squark.ai.
Competition-driven learning on the RSNA challenge. 3D CT volumes, class imbalance, and competition-grade ensembling — Kaggle as a lab bench, not a trophy case.
Designer-developer bridge project — generative palettes, perceptual constraints, and tools that speak both Figma and PyTorch. RISE 2019 finalist. Still iterating.
AFFILIATED WITH · COLLABORATING WITH
Reach out for collabs — events, hackathons, joint research.
§ 04 — Charter
We do not build features.
We build the things that make features look obvious in retrospect.
Taste is the only moat that doesn't decay.