GSOC 2026 Project #33 : University of Wisconsin-Madison - AStats: an agentic-AI approach to applied statistical practitioner workflows

GSoC
42

I added some comments.. I dont have to send you a DM first to unlock it.. perhaps you have to post first in order to be able to send DMs, no idea.

43

Hi, following up on my earlier question and prototype.

Based on your suggestion to explore with actual datasets and extend the idea further, I worked on building a more complete structure-aware statistical pipeline.

The current prototype includes:

  • basic data profiling

  • structure detection (distinguishing grouping vs repeated measures)

  • an ambiguity detection step to avoid premature decisions

  • decision logic for selecting the statistical test

  • an LLM-style reasoning layer to explain the choice

While working on this, I noticed that simple heuristics can easily misclassify continuous variables or grouping columns as subject-level structure, which leads to incorrect repeated-measures detection. I refined the approach by adding stricter constraints for subject identification and validating structure across groups, which improved the correctness of the pipeline.

I also added a basic ambiguity handling step so that the system can defer decisions and indicate when user clarification may be needed, instead of making incorrect assumptions.

I tested this on both simulated repeated-measures data and a real dataset (Iris), and it correctly distinguishes between dependent and independent settings in these cases.

Prototype:
https://github.com/pragati-0208/astats-structure-experiments

I’m planning to extend this further by incorporating richer profiling signals and exploring LLM-assisted decision-making more deeply.

Would appreciate any feedback on whether this direction aligns well with the intended workflow design.

Thanks!

44

Hi @suresh.krishna,

I have started building a working prototype for AStats and wanted to share my progress. I have sent a detailed email to the mentor (suresh.krishna@mcgill.ca) covering what I have built, my understanding of the project direction, and a few questions before I finalize my proposal.

GitHub repo: GitHub - sarthak-here/astats: Agentic AI system for applied statistical workflows — GSoC 2026 Project #33 (INCF) · GitHub
Happy to discuss anything here on the thread as well.

Thank you!

45

Thank you for your interest and work.

We can provide feedback on1 version of the full draft, as mentioned above. – you can send me a link here and provide access only to my email account.

Question 3 is answered in the project announcement (yes). Your approach in 1 is fine and for 2, depends on what you think is feasible.

Good luck !

46
48

Thank You !! @suresh.krishna for accepting my request in Astats Community .

49

Hi Suresh,

Thank you for the feedback — I’ve incorporated everything into version 2 and will be submitting on the GSoC site before the deadline.

Updates made:

  • Expanded the GPT-4.1 benchmark with a category-by-category breakdown

  • Validated on 6 real published datasets (sleepstudy, iris, ToothGrowth, PlantGrowth, ChickWeight, congruency RT) — 15 total

  • Added Claude, Codex, Ollama, and offline backends via --llm flag

  • Added practitioner user journey diagram and terminal screenshots

Prototype: https://github.com/ShreeniG99/AStats-Agentic-AI-for-Applied-Statistical-Practitioner-Workflows

Proposal: https://docs.google.com/document/d/1ESWjn2MXSeTjWH9BOV9hQuh6OG_qiac5/edit?usp=drive_link

Regards, Shreenidhi Gopalakrishnan

1 Like
50

Hi @suresh.krishna ,

In real world , Many types of data exist … Tabular data ,text data , Imge based data , voice datasets ,signals based dataset . Before building any AI system first of all we have to look at the DATA … What type of data we are going to feed into our sytem as input then we can design and implement it in better way . So while building Astats what type of data ingestion we are mainly targetting ?

Thanks !!

Regards
Areeba Tahir

51

In neuroscience , I have seen data in EEG ,FMRI ,IMAGES … . This is an edge case in building Astats . So What about designing a system that accepts raw EEG/fMRI/eye-tracking data and automatically extracts statistical features, then feeds them into the same structure aware pipeline ? @suresh.krishna

52

Hi, I’ve submitted my proposal for AStats and wanted to share what I’m planning to build. I’d really appreciate your feedback.

The idea is to make statistical analysis simple and accessible. Users can provide their dataset and ask questions in plain English, and the system will:

  • Understand the data type

  • Select the appropriate statistical test

  • Run the analysis

  • Clearly explain the reasoning behind each decision

It will also handle:

  • Normality checks (and switch methods if assumptions fail)

  • Multiple comparisons (apply FDR correction)

  • Low sample sizes (flag underpowered studies)

The goal is full transparency — no black-box decisions.

I’ve previously built a similar system for Natural Language to SQL. Users can input queries in plain English, and the system generates SQL, connects to the database, executes it, and returns results.
Example: “Show me top 10 movies by rent” → automatically fetches and displays results.

You can check it out here: https://nl2sqlexplorer.netlify.app
Blog: https://medium.com/@rj.yogeshwari/transform-natural-language-into-sql-queries-with-nl2sql-explorer-242ea659c86c

Would love to hear your thoughts or suggestions!

53

Hi,

Thank you for proposing the AStats project. I am incredibly interested in this idea because it perfectly matches my passion for building agentic AI workflows and integrating LLMs into practical tools.

My name is Ayaan Ahmed Khan, and I am a Computer Science undergraduate specializing in Artificial Intelligence and Full-Stack development at COMSATS University Islamabad, Pakistan. I have a lot of experience building AI-driven applications from scratch. Recently, I built Padhlo AI, an educational tool using the Gemini API, where I had to design strict prompts to ensure the AI generated accurate outputs without hallucinating. I also have a strong background in Python data science, having built both supervised and unsupervised machine learning models using NumPy, pandas, and scikit-learn.

While I am not yet deeply experienced with advanced HPC environments like Slurm, I am a very fast learner. I have already started researching how to securely execute Python code in sandbox environments for these kinds of AI agents. My background as a Frontend Team Lead and my experience learning new tech stacks quickly gives me the confidence that I can adapt to any new tools this project requires.

Since the GSoC deadline is today, I am currently submitting my formal proposal for AStats! However, I would love to hear any thoughts or advice you have, especially regarding which specific open-weight models the team is most interested in testing for the local environment.

I am highly motivated to work on this project and am ready to dedicate my full summer to making AStats a reality.

Best regards,

Ayaan Ahmed Khan