Gsoc mvp

[ad23b1012]

Overview

This PR introduces the completed Phase 1 Minimum Viable Product (MVP) for the AStats project, designed for the UW-Madison GSoC 2026 proposal. It establishes a fully functional, end-to-end agentic AI framework capable of exploring datasets, generating statistical visualizations, executing Python code, and compiling professional analytical reports — validated against real-world scientific datasets.

🚀 What Was Accomplished (Phase 1)

We built the entire modular framework entirely from scratch, ensuring best practices in software architecture and AI integration:

1. Core Agentic Architecture

• Plan-Execute-Reflect Loop: Implemented a robust BaseAgent class that autonomously guides the LLM through a structured reasoning cycle with iterative self-correction on execution errors.

• Specialist Agents: Created domain-specific agents (EDAAgent, HypothesisAgent, RegressionAgent, TimeSeriesAgent) for targeted statistical methodologies.

• Workflow Orchestrator: Developed an intelligent router that analyzes user queries to automatically select the best agent, with configurable autonomy levels (full-auto, semi-auto, step-by-step).

2. Multi-Provider LLM Abstraction

• Built a flexible BaseLLMProvider abstraction layer to prevent vendor lock-in.

• Integrated 4 providers:

  • Google Gemini 2.5 Flash — Primary free-tier default

  • Groq (Llama 3 70B / Mixtral) — Blazing-fast open-weight inference

  • Anthropic Claude 3.5 — Sonnet, Opus, and Haiku support

  • OpenAI / Codex — GPT-4o and legacy Codex support

• All providers include streaming, structured function calling, and automatic retry with exponential backoff.

3. Data Engine & Tool Registry

• Intelligent Tooling: Provided agents with sandboxed execution environments via the ToolRegistry (create_plot, describe_data, run_code, run_statistical_test, fit_model).

• Auto-Profiler: Built a high-speed profiling engine to ingest datasets (CSV, JSON, Parquet, Excel, Stata, SPSS, Feather, etc.) and generate LLM-readable statistical summaries.

• Smart Visualization Engine: Auto-selects distributions, correlation heatmaps, box plots, and categorical charts based on data profile.

4. Real-World Dataset Validation

Validated the full pipeline end-to-end on standard scientific datasets:

Dataset	Rows	Task	What It Demonstrates
Fisher's Iris	150	Exploratory Data Analysis	Auto-profiling, species grouping, correlation discovery
Diabetes	442	Regression Modeling	OLS fitting, VIF multicollinearity check, Ridge/Lasso regularization
Titanic	100	Hypothesis Testing	Chi-Square, independent t-test, Mann-Whitney U, One-Way ANOVA

Each dataset has a fully documented example script in examples/.

5. Statistical Methodology Documentation

• Created METHODOLOGY.md documenting the rigorous statistical workflows: EDA (normality screening, outlier detection, bivariate associations), Hypothesis Testing (parametric & non-parametric with effect sizes), and Regression (OLS diagnostics, Durbin-Watson, Breusch-Pagan, VIF, regularization).

6. CLI, UX & Reporting

• Interactive CLI: Built a Click-based CLI with subcommands (explore, analyze, profile, config).

• HTML/Markdown Reports: Implemented automated, beautifully formatted HTML report generation using Jinja2, with Base64-embedded Seaborn/Matplotlib plots for zero-dependency sharing.

🔮 What's Next (Phase 2 & Beyond)

Now that the core framework is mathematically and architecturally sound, the next steps for the project entail:

1. Local Open-Weight Integration: Wiring up local inference servers (e.g., Ollama or vLLM) so the analysis can be run entirely air-gapped without API keys.

2. Fine-Tuning Pipelines: Researching and building dataset generation pipelines to fine-tune a smaller LLM specifically on data practitioner reasoning.

3. Advanced Evaluators: Implementing "critic" agents to double-check statistical code and assumptions before execution.

4. Workflow Templates: Developing reusable, recipe-based workflow templates for common statistical analyses.

🧪 Testing Notes

• Tested end-to-end on sample_sales.csv, iris.csv, diabetes.csv, and titanic.csv.

• All 3 example scripts (eda_iris_example.py, regression_diabetes_example.py, hypothesis_titanic_example.py) execute successfully and produce correct statistical outputs.

• Verified environment variable injection (.env) for seamless API key configuration across all 4 providers.

• All generated plots are verified to embed natively via Base64 in output HTML reports.

[Mustafa0216]

i have uploaded my work in my repo mentors please check my work there