Cursor Boston × PyData Data Science Hack

This is a sharp and technically ambitious notebook: it adapts speculative decoding from LLM serving to robot diffusion policies and backs the claim with latency, acceptance-rate, MSE, and cross-episode validation. The narrative is clear, the 10 Hz deadline makes the result actionable, and the interactive tau/Pareto exploration uses Marimo well. The main caveat is that the methodology is complex enough that a reader has to trust several implementation details, but the evidence presented is unusually strong for a single-evening submission.

Refined tie-break score: this notebook delivers the strongest winner-eligible 8-range result because it turns the UCI Bank Marketing data into a crisp, operational recommendation: the bank underused the 31-90 day recontact window where conversion was strongest while over-dialing weaker late windows. It adds strong visual contrasts, checks the obvious seasonality confound within May, estimates missed opportunity, and separately analyzes dial-count exhaustion. It is less technically ambitious than the 9/10 entries, but it has the clearest actionable insight among the tied 8s.

Refined tie-break score: this notebook builds a thoughtful composite measure of transit privilege by combining rapid-transit access, live alert exposure, and chronic Blue Book wait-time reliability. The geospatial joins, ethical framing, sensitivity controls, and factor breakdown make it more rigorous than a simple map. It ranks just below the top 8-range entry because the live-alert component is time-sensitive and the composite weighting remains exploratory, so the finding is useful but less cleanly decisive.

Refined tie-break score: this is a polished, well-explained exoplanet explorer that combines discovery-history storytelling with interactive Earth-like candidate filtering and transparent ranking. The controls and candidate cards are strong for public exploration. It ranks below the higher 8-range entries because the main result is broader educational exploration rather than a single sharply defended quantitative claim.

Refined tie-break score: this notebook is a polished Marimo narrative around the Fulton Gap, with careful sample cuts, effective visualizations, and an interactive irradiation filter that demonstrates the expected radius-valley shift. It is elegant and accessible, but among the tied 8s it is least novel because it primarily recreates a known published result rather than surfacing a new actionable or original finding.

The notebook builds a practical MBTA reliability baseline from heavy-rail travel-time data, with route-pair variability metrics, unstable segment rankings, and anomaly-day detection. The operational framing is clear and the Polars-based workflow is credible for larger transit files. It falls short of the top tier because it depends on local CSV availability and the key insight is more of a reliability tooling foundation than one surprising, tightly defended finding.

This final notebook has a clear early-triage story for Spike RBD mutations and turns ProteinGym DMS measurements into flexible-site, brittle-site, and high-binding mutation rankings. The narrative framing is strong and the heatmap/watchlist format is useful for quickly prioritizing variants. It is less rigorous than the top entries because it mostly summarizes single-assay fitness rather than validating whether the watchlist predicts real surveillance or immune-escape risk, but it is a solid and relevant analysis.

The notebook provides a useful interactive geospatial explorer for Boston liquor licenses, with clustered markers, density heatmap, and time-of-day filtering. It is polished as a map interface and uses city open data effectively, but the insight is mostly descriptive concentration/exploration rather than a surprising or rigorously defended finding. The work is solid, especially visually, but less analytical than the top submissions.

The notebook provides a polished narrative tour of Boston economic indicators, with useful attention to missingness/zero-filled legacy fields and multi-series context around jobs, mobility, hospitality, housing, and development. It is visually and rhetorically engaging, but the core insight is diffuse: it reads more like an exploratory dashboard than one sharply defended economic indicator or surprising finding. The analysis is solid enough to be useful, but not as focused or rigorous as the strongest entries.

The notebook builds a useful RFM customer segmentation dashboard on UCI Online Retail, with cleaned transaction data, segment labels, and practical marketing playbooks. It is business-friendly and interactive, but the core RFM approach is standard and the AI recommendations are prewritten playbook copy rather than model-derived evidence. It is a solid applied dashboard rather than a surprising analytical insight.

The notebook has a coherent idea: compare MBTA route headway unpredictability and relate it to weather exposure, with a concrete Green Line vs Blue Line framing. The live CV calculation and chart make the finding approachable, but the analysis is fragile because it depends on a live API snapshot and a local weather CSV that may be absent, and the weather connection is more suggestive than rigorously supported. It is a reasonable exploratory submission but not a fully defended insight.

The notebook loads the UCI Bank Marketing data, gives summary statistics, and explores whether more campaign contacts are associated with customer balance. It is honest about missing customer/loan timing limitations, but the analysis is narrow, mostly descriptive, and does not establish a strong or actionable insight. The result is a small exploratory check rather than a compelling hackathon finding.

The PR has a metadata file but the submitted `submission.py` is empty, so there is no notebook, analysis, evidence, or insight to evaluate. This is therefore a broken/empty submission under the rubric.

This is an unusually ambitious and well-framed public-data atlas: it combines GTEx tissue expression, ProteinGym coverage, ClinVar, PubMed, Open Targets, and other evidence layers to ask which PCD mechanisms are visible or underrepresented in public molecular data. The notebook is careful about provenance, limitations, and clinical boundaries while still producing a useful evidence-gap map. It is broad rather than a single narrow statistical punchline, but the scope, rigor, and storytelling make it one of the strongest submissions.

The notebook extracts a sharp story from the UCI Bank Marketing data: conversion rises dramatically across the inferred 2008-2010 campaign timeline, and the analysis connects that lift to warmer targeting and prior-success signals while excluding post-call duration from predictive modeling. It is well structured, uses sanity-check assertions, and balances descriptive trend analysis with realistic modeling. The main caveat is that the inferred year reconstruction depends on row ordering, but the author is explicit about that assumption.

The notebook is a polished and imaginative exoplanet mission-planner interface with NASA Exoplanet Archive integration, Earth-similarity scoring, interactive controls, Plotly candidate ranking, and a clear mission-mode framing for nearer-star prioritization. The metadata update now gives the submission a complete package. It is still more of an exploratory decision-support tool than a rigorously validated scientific analysis, but the presentation, interactivity, and coherent scoring workflow make it one of the stronger submissions.

This replacement notebook improves the original Second Earth Scout concept with a clearer habitability-intelligence narrative, NASA Exoplanet Archive data, Earth-similarity and habitable-zone scoring, clustering, anomaly detection, and Random Forest ranking. The dark-horse framing is engaging and the analysis is more polished than the first version, but the model still relies on synthetic labels derived from handcrafted habitability scores, so the evidence is more exploratory than independently validated. Overall it is a strong interactive exploration with moderate methodological rigor.

The notebook presents a clear campaign timing insight from the UCI Bank Marketing data: conversion varies sharply by job segment and month, with retired and student customers performing especially well in March, September, and October. It also includes useful call-fatigue analysis and correctly warns about duration leakage for predictive pipelines. The analysis is focused and actionable, though it is mostly descriptive segmentation rather than a deeper causal or validated modeling study.

The notebook is ambitious and user-facing: it combines MBTA live data, equity scorecards, ghost-bus heuristics, multilingual copy, RAG-style policy snippets, and advocacy outputs into a civic dashboard. The scope and practical framing are strong, but some demographics/policy inputs are illustrative or hardcoded and the LLM features depend on optional API keys, so the evidence is less rigorous than the interface suggests. It is a compelling prototype with meaningful civic value, but not a fully validated transit equity analysis.

The notebook asks a strong benchmark question: where do all ProteinGym models fail collectively, rather than which model wins on average. Joining model-level Spearman summaries to assay metadata and linking hard proteins to AlphaFold structure confidence is an interesting and useful failure-mode analysis. It is held back by a few unfinished interpretation notes and a relatively light final synthesis, but the core idea and implementation are solid.

This is a coherent Marimo notebook with a strong technical through-line: it connects a classic UCI robot wall-following dataset to modern LeRobot episode structure, compares logistic regression and MLP baselines, and shows how lagged observations approximate short-term memory. The framing is clear and the code is compact. It loses points because the PR packaging was messy, the original metadata was empty, and the analysis is mostly a compact model comparison/schema sketch rather than a deeper diagnostic study with visualizations or extensive interpretation.

The notebook is a polished self-contained NLP exploration that maps 24 historical-thinker persona prompts with TF-IDF, cosine similarity, PCA, KMeans, and distinctive-term tables. It is creative and readable, but the corpus is hand-authored and small, so the analysis is more of a thoughtful visualization exercise than a strongly evidenced external-data insight. The presentation is strong, while the empirical stakes are modest.

The submission builds a clear Boston crime and safety analysis with district-level risk scoring, heatmaps, crash concern overlays, and a concrete finding that B2, B3, and C11 are highest risk while A1 and A7 are safer. The interactive map is useful and the civic-data framing is strong. It loses points because the notebook is an exported Jupyter script rather than a polished Marimo notebook, installs packages at runtime, relies on browser/camera gesture controls that are not central to the analysis, and uses a simple severity heuristic without much validation or uncertainty analysis.

The notebook gives a broad and practical Online Retail analysis: cleaning, time/geography/product/customer EDA, RFM segmentation, revenue concentration, and churn-style simulations. It is useful as a business dashboard and covers the dataset well, but the core finding is a fairly standard concentration/RFM story rather than a surprising single insight. The work is solid, but the narrative and rigor are more walkthrough-like than exceptional.

The notebook has an interesting descriptive geospatial question: whether nighttime Bluebikes trip ends are closer to Boston streetlights than daytime trip ends, with nearest-neighbor distances and map-style visualizations. The one-line takeaway is clear and it acknowledges confounders, but the submission depends on local data files not shown in the PR and originally had a syntax typo, so reproducibility and polish are weak. It is a reasonable exploratory analysis but not a fully robust insight.

The notebook explores MBTA GTFS and reliability data with route-type counts, bus shape visualizations, and bus reliability mapping. It has a relevant civic-data theme, but the analysis is mostly exploratory infrastructure plotting, includes a weak opening/metadata package, and does not clearly answer the stated on-time reliability question with a crisp finding. It is a useful start, but less polished and less insight-focused than stronger submissions.

The notebook outlines a useful interactive dashboard for public salary/employee earnings data, including department, title, ZIP, and map-based breakdowns over 2011-2025. However, it depends on local data files that are not included in the PR, starts with placeholder cells, and the analysis is mostly descriptive distribution plotting rather than a sharp finding. The dashboard idea is valuable but the submitted package is only partially reproducible.