Ship Better Data Science Projects with AI

I am building a [classification/regression] model. Here is my situation: dataset size is [number] rows and [number] features. The target variable is [describe — continuous, binary, multiclass]. Class balance: [balanced/imbalanced — give ratio]. Feature types: [number] numerical, [number] categorical, [number] text. Latency requirement: [real-time / batch]. Interpretability requirement: [must explain predictions / black box is fine]. Recommend the top 3 algorithms I should try, ranked by expected fit for this scenario. For each algorithm: (1) explain why it suits my constraints, (2) list the key hyperparameters to tune and reasonable ranges, (3) flag potential pitfalls for my specific data characteristics, and (4) provide the scikit-learn or XGBoost code to train a baseline version with sensible defaults.

I am tuning a [model name] for [task description]. Current performance: [metric] = [value]. The hyperparameters available are: [list hyperparameters]. My compute budget allows [number] total training runs. Design a tuning strategy: (1) recommend which hyperparameters to tune first based on typical impact (rank by importance), (2) suggest the search method (grid search, random search, Bayesian optimization) and justify the choice for my compute budget, (3) define the search space for each hyperparameter with specific ranges, (4) provide the complete Python code using Optuna or scikit-learn for the search, (5) explain early stopping criteria to avoid wasting compute on bad combinations, and (6) describe how to analyze the tuning results to understand which hyperparameters matter most.

I need to evaluate my model properly. My data: [describe — size, time component, groups/clusters, class balance]. The model is a [model type] predicting [target]. Design the right cross-validation strategy: (1) recommend the CV method (k-fold, stratified k-fold, group k-fold, time series split, nested CV) and explain why it fits my data structure, (2) explain what would go wrong if I used simple k-fold instead, (3) specify the number of folds and any shuffle settings, (4) provide complete Python code implementing the CV with proper metric reporting (mean, std, per-fold results), (5) include a visualization of the CV splits so I can verify they look correct, and (6) explain how to use the CV results to estimate real-world performance with appropriate confidence bounds.

I have a binary classification problem with severe class imbalance. Positive class: [number] samples ([percentage]%). Negative class: [number] samples ([percentage]%). Model: [current model]. Current metrics: precision = [X], recall = [X], F1 = [X], AUC = [X]. Create a comprehensive plan: (1) recommend 3 resampling techniques (SMOTE, ADASYN, random undersampling, etc.) and explain which fits my situation best, (2) suggest 3 algorithmic approaches (class weights, threshold tuning, cost-sensitive learning), (3) provide Python code implementing the top recommendation with proper cross-validation (resample inside the fold, not before), (4) explain which evaluation metrics to use and which to ignore for imbalanced data, (5) show how to find the optimal classification threshold using precision-recall curves, and (6) warn about common mistakes that make imbalanced problems look solved when they are not.

I have trained [number] different models for [task]: [list model names]. I need to decide which one to deploy. Provide a rigorous model comparison framework: (1) list the metrics I should compare (accuracy, precision, recall, F1, AUC, log loss, calibration — explain which matter most for [business context]), (2) write Python code to run all models through the same cross-validation with a paired comparison, (3) run McNemar's test or a corrected paired t-test to determine if performance differences are statistically significant, (4) create a radar chart comparing all models across multiple metrics, (5) add a complexity comparison — training time, inference time, model size, interpretability score, (6) produce a final recommendation table with a weighted scoring system based on my priorities: [rank importance of accuracy, speed, interpretability, maintainability].

I need to create a data dashboard for [audience — executives, product team, clients]. The key metrics are: [list 5-8 KPIs with brief descriptions]. The data updates [daily/weekly/monthly]. Design the dashboard: (1) recommend the layout — which metrics get top placement and why, (2) specify the chart type for each metric (line, bar, gauge, number card, etc.) with justification, (3) define the color scheme and visual hierarchy, (4) include comparison context for each metric (vs last period, vs target, vs benchmark), (5) suggest 2-3 drill-down views for power users, and (6) provide complete Python code using Plotly Dash or Streamlit to build a working prototype. The dashboard should tell a story, not just display numbers.

I need to create [number] plots for a [report/presentation/paper]. The data shows: [describe the key finding each plot should communicate]. For each plot: (1) recommend the best chart type to communicate that specific finding, (2) provide complete matplotlib or seaborn code with publication-quality formatting — proper font sizes (12pt+ for labels), high DPI (300), clean white background, minimal gridlines, (3) include informative titles and axis labels that state the finding, not just the variable names, (4) add annotations highlighting the key takeaway, (5) use a colorblind-friendly palette, and (6) export as both PNG and SVG. Each plot should be understandable without reading surrounding text.

I have a dataset about [topic] with columns: [list columns]. Create a Jupyter notebook with interactive visualizations using Plotly that lets me explore the data dynamically: (1) an interactive scatter plot with dropdown selectors for x-axis, y-axis, and color-by columns, (2) a filterable histogram with a slider for bin size, (3) an interactive correlation explorer where clicking a cell in the heatmap shows the underlying scatter plot, (4) a parallel coordinates plot for comparing [3-5 numerical features] across [categorical grouping], and (5) a time series viewer with range slider if there is a date column. Include ipywidgets for any controls. Add markdown cells explaining what to look for in each visualization.

I have data with geographic components: [describe — lat/long coordinates, country names, state/province, zip codes, city names]. The metric I want to visualize is [describe metric]. Total rows: [number]. Create a Python script using Folium or Plotly that: (1) creates a choropleth map colored by [metric] at the [country/state/county] level, (2) adds interactive tooltips showing the location name and metric value on hover, (3) includes a heatmap layer for point data if I have lat/long coordinates, (4) adds a legend with meaningful color breakpoints (not arbitrary quantiles), (5) implements clustering for dense point data to avoid overplotting, and (6) exports both an interactive HTML version and a static image for reports. Use a sequential color scale appropriate for [metric type].

I have a trained [classification/regression] model. Predictions and actuals: [Describe or paste sample of y_true, y_pred, y_prob] Create a complete model performance visualization suite: (1) confusion matrix with counts and percentages, (2) ROC curve with AUC score annotated (for classification), (3) precision-recall curve with average precision (for classification), (4) calibration plot comparing predicted probabilities to actual frequencies, (5) residual plots — predicted vs actual, residual distribution, residual vs each feature (for regression), (6) feature importance plot — bar chart of top 20 features, and (7) SHAP summary plot showing feature impact on individual predictions. Use a consistent visual style across all plots and arrange them in a 2x4 grid for a single-page overview.

I completed a data science analysis and need to present findings to [audience — C-suite, product manager, non-technical client]. Here are my technical findings: [Paste your technical findings, including model performance, key statistics, and conclusions] Rewrite this as a one-page executive summary that: (1) starts with the business question and the bottom-line answer in the first sentence, (2) translates all statistical results into business impact (dollars, percentages, customer counts — not p-values or R-squared), (3) includes exactly 3 key findings with one supporting data point each, (4) states the recommended action clearly, (5) acknowledges limitations in one sentence without undermining confidence, and (6) ends with a specific next step and timeline. Use no jargon. A smart executive with no data science background should understand every word.

I need to document a data science project so my team can understand, reproduce, and extend it. Project: [describe project]. Model: [describe model]. Data sources: [list]. Write a documentation template that covers: (1) project overview — problem statement, stakeholders, success criteria, (2) data dictionary — table with every column, its type, description, source, and any transformations applied, (3) methodology — model selection rationale, feature engineering decisions, validation approach, (4) results — key metrics with confidence intervals, comparison to baseline, (5) deployment notes — how to retrain, input/output formats, latency expectations, (6) known limitations and failure modes, (7) monitoring plan — what metrics to track post-deployment and alert thresholds, and (8) changelog — version history with what changed and why.

I am presenting a data science project to [stakeholder group]. My analysis shows [summarize key finding]. The recommended action is [action]. Anticipate the 10 toughest questions stakeholders will ask, grouped into: (1) methodology challenges — "Why did you use this approach?", "How confident are you?", (2) data quality concerns — "Is the data reliable?", "What about missing data?", (3) business impact questions — "What is the ROI?", "How does this compare to the status quo?", (4) implementation concerns — "How long to implement?", "What are the risks?", and (5) political/organizational challenges — "Does this affect my team?", "Why should we change?" For each question, provide a concise, confident answer with one supporting data point. Include a "bridge" phrase to redirect hostile questions back to the key finding.

I need to create a recurring [daily/weekly/monthly] data report that runs automatically. The report should cover: [list 4-6 metrics or sections]. Data sources: [describe — SQL database, CSV files, API]. Audience: [who reads it]. Write a complete Python script that: (1) connects to [data source] and pulls the required data, (2) calculates all metrics with period-over-period comparisons, (3) generates visualizations using matplotlib, (4) formats everything into a clean HTML email or PDF using Jinja2 templates, (5) highlights any metrics that breach alert thresholds in red, (6) includes an automatically generated commentary section for significant changes, and (7) sends the report via email using smtplib or saves to a shared folder. Include error handling and logging so I know when the report fails.

I have the following data insights from my analysis: [List 5-7 findings with their supporting metrics] Help me weave these into a compelling data story for [audience]. Structure it as: (1) the hook — start with the most surprising or high-stakes finding to grab attention, (2) the context — what was happening before, what question we set out to answer, (3) the journey — walk through 3-4 findings in a logical sequence where each builds on the previous one, (4) the climax — the key insight that changes how we should think or act, (5) the resolution — the specific recommendation and expected outcome, and (6) the call to action — what needs to happen next and by when. For each section, suggest the supporting chart or visual. Write it as a presentation script I can deliver in [10/15/20] minutes.

I need to write a SQL query against this schema: [Describe tables, columns, primary/foreign keys, or paste CREATE TABLE statements] The business question is: [describe what you need to answer, e.g., "What is the monthly revenue per customer segment for the last 12 months, including customers who made no purchases"]. Write the SQL query with: (1) proper JOIN types and explain why each is LEFT/INNER/FULL, (2) window functions where they simplify the logic, (3) CTEs for readability if the query is complex, (4) handling of NULL values and edge cases, (5) performance considerations — suggest indexes if appropriate, and (6) comments explaining each major section. Test the logic by describing what each CTE or subquery produces. Optimize for readability first, then performance.

Here is my current pandas code that works but is slow or messy: [Paste your pandas code] The DataFrame has [number] rows. Optimize this code: (1) identify operations that can be vectorized instead of using apply or loops, (2) replace any iterrows with vectorized alternatives, (3) chain operations using method chaining for readability, (4) use categorical dtypes for string columns with low cardinality, (5) suggest where .query() or .eval() could improve readability, (6) benchmark the original vs optimized version with estimated speedup, and (7) add type hints and docstrings. Explain each optimization and why it is faster. Preserve the exact same output — do not change the logic, only the implementation.

I am building a data pipeline in Python that: [describe pipeline — reads from source, transforms, loads to destination]. Write a robust version with proper error handling: (1) add try/except blocks with specific exception types (not bare except), (2) implement retry logic with exponential backoff for API calls and database connections, (3) add data validation checkpoints — row counts, schema checks, value range checks — after each step, (4) implement logging using the logging module with appropriate levels (INFO for progress, WARNING for recoverable issues, ERROR for failures), (5) add a dead-letter queue or error table for rows that fail validation, (6) implement idempotency so the pipeline can be safely rerun without duplicating data, and (7) send an alert (email, Slack, or log) on failure with the error details and the step that failed.