@Roaa-838 thanks for the interest. Please treat this as your prototype/POC and describe it along your vision in your proposal. And please go through the other messages here and on scicommons as well. All the best, and we can provide feedback with 1 version of your proposal. Regarding yor questions, for question 1, you can choose and for question, subprocesses.
please check that: https://drive.google.com/file/d/1h7JcUH8hUsCnQOhcgbfEOETZakxunCYL/view?usp=sharing
Hi Suresh,
Thanks for the clarification.
I implemented a small prototype to explore how early-stage structure understanding affects statistical outcomes.
In a simulated repeated-measures dataset:
-
Treating samples as independent → non-significant result
-
Using the correct paired test → strongly significant result
This shows that the failure happens before test selection, at the stage of structure interpretation.
I then added a simple heuristic layer to detect grouping columns and flag possible repeated-measures structure.
Prototype:
https://github.com/pragati-0208/astats-structure-experiments
This is still very preliminary, but it helped me think about how a structure-aware profiling stage could reduce early errors.
I’ll try to extend this further into a more coherent pipeline.
Would appreciate any feedback if this direction seems reasonable.
Thanks!
1 Like
@suresh.krishna Hi Suresh,
Here is my proposal, please take a look on it when you are available.
I am looking forward to implement the propotype after getting your feedback.
Thank you.
Proposal: https://docs.google.com/document/d/108PnIub0koosSTsD7nBWpgvwMEzFgnSVGtaqlx7BYbM/edit?tab=t.0
1 Like
@Ducanh_Dangtrung - your basic structure is fine.. you should flesh it out and provide more detail with your vision and how you will execute it, and then you can send it to me via DM. We can provide detailed feedback on one version of your proposal.
I suggest that you think carefully about how such a framework may look like in a usable functional form, and how you would go about setting it up and performing the first steps towards it.. and then see how much of that can be done within the GSoC period. Keeping an eye on the statistical workflow that practitioners use, and how you might go about using an LLM to aid you in real statistical data analysis, will all serve as useful pointers.
And do use the INCF template.
All the best !
Hello Jonathan, Yohai-Eliel, and Suresh,
I’m Shreenidhi Gopalakrishnan, a 4th-semester B.Tech student in AI & Data Science at Amrita Vishwa Vidyapeetham, with a Minor in Medical Engineering.
Reading through this thread, I noticed that most contributors have focused on the NL query layer and the assumption-checking layer. I’d like to approach AStats from a different angle — the human-in-the-loop feedback architecture — which I think is the hardest and least-addressed part of making agentic statistical workflows genuinely trustworthy in practice.
My background that directly shapes this perspective: I built a clinical decision support application for adverse drug reaction prediction in Hepatitis B antiviral therapy (LightGBM + FDA FAERS data, deployed with Streamlit), where the core challenge was exactly what AStats faces — an automated ML pipeline whose outputs clinicians need to interrogate, override, and trust. I integrated SHAP-based explainability specifically to give practitioners a way to see why a prediction was made and push back on it. That experience gave me a strong sense of where agentic systems fail practitioners: not in the test selection logic, but in the moment where the agent’s confidence and the human’s domain knowledge conflict.
I’ve also built multi-agent LangGraph pipelines (a Stage-Aware Career Co-Pilot at the IETE AI-Verse hackathon) and RAG systems with BioBERT/BioGPT fine-tuned with LoRA (BioChat AI), so the agentic infrastructure side is familiar territory for me.
The specific design question I’m thinking about is that “When AStats produces a statistical recommendation and the practitioner disagrees or overrides it, how should that correction propagate? As a one-time override, as a persistent preference that shapes future recommendations, or as a signal for fine-tuning the underlying model?”
I’ve requested to join the AStats community at alphatest.scicommons.org and am beginning to sketch a prototype for a practitioner feedback layer.
Looking forward to your guidance.
Best regards,
Shreenidhi Gopalakrishnan
1 Like
You can use A or B depending on user preference and C if user agrees.. in general, for such questions, you can make any choices you wish. The proposal itself should be at a broader level. Good luck ! We can offer comments on 1 draft version.
Yes, these are the sort of initial experiments that should guide your explorations and proposal. All the best @Pragati0208
Hi Suresh,
I have prepared a draft proposal and would be grateful for the one version of feedback you mentioned. https://drive.google.com/file/d/1q2xxuUqqV87INNiVLLXzBVjISUmFvClp/view?usp=sharing
You can leave inline comments directly on the document. I will incorporate all feedback before final submission through the GSoC portal.
Thank you for your time!
1 Like
Hi @suresh.krishna,
Thank you for the guidance. Based on your suggestion, I’m approaching the proposal from a broader perspective by focusing on designing an agent-based statistical workflow system rather than a fixed pipeline.
My idea is to build an agentic framework where different components (EDA, modeling, inference, and explanation) act as specialized agents, coordinated by a decision layer that can choose between multiple approaches (A/B/C) based on data characteristics and user preferences.
I’m also considering adding features like explainability and reproducible workflows to make the system more useful for practitioners.
I’ll validate these ideas through initial experiments before finalizing the proposal.
Looking forward to your feedback!
Thank you!
1 Like
Hi Suresh
I’ve been reading through the AStats idea and what stood out to me most is that it is not just about building an agent, but about shaping a workflow that feels reliable, interpretable, and useful for real statistical practice. I really like the direction of combining structured analysis with agentic support, because it feels much closer to how practitioners actually work than a fully open ended agent.
What interests me most is how the system can balance flexibility with reproducibility starting from a clear statistical workflow, then adding agent support where it genuinely improves exploration, I would be grateful to learn what you think should be the first principle in designing this kind of system should the priority be workflow structure, decision reliability, or adaptability?
1 Like
All 3. And welcome !
Great and welcome ! Looking forward to your proposal.
Thanks @suresh.krishna that makes sense.
I’ll try to explore this by starting with a small structured workflow where the system prioritizes decision reliability and reproducibility, and then gradually introduce adaptability on top of it to see how it impacts the outcomes.
I’ll share some initial experiments soon to get feedback.
1 Like
@NssGourav - given the very short time left for proposal submission, at this point, we can offer comments on 1 version of the proposal. Once the coding period starts, we will see who is still around (GSoC intern(s) and/or volunteers) and then we can create a plan to work together. All the best for your proposal.
1 Like
Hi Suresh,
I have prepared the draft proposal and would be grateful for your feedback.
Thank you
Hi everyone,
I’m Abhishek Buddiga, and I’m excited to submit my proposal for GSoC 2026 – AStats: An Agentic-AI Approach to Applied Statistical Practitioner Workflows.
I’ve been closely following the discussions around this project and found the focus on workflow design, human-guided analysis, and reliable statistical reasoning with agentic AI particularly compelling. Over the past few days, I’ve explored these ideas in depth and also built a small exploratory prototype to better understand the design space and challenges.
In my proposal, I focus on:
-
Designing structured, recipe-driven statistical workflows
-
Ensuring reliability via tool-grounded execution
-
Incorporating human-in-the-loop interaction
-
Supporting open-weight models for reproducibility and control
-
Defining best practices for agentic statistical systems, not just building a tool
I would greatly appreciate your feedback on the approach, especially regarding:
-
Workflow design choices
-
Evaluation strategy
-
Balance between automation and practitioner control
Proposal: https://docs.google.com/document/
MVP PR (Exploratory Prototype): https://github.com/ad23b1012/AStats/pull/new/gsoc-mvp
Thank you for your time and guidance I’m really looking forward to contributing to AStats and collaborating with the mentors and community.
Best regards,
Abhishek Buddiga
Hi @suresh.krishna ,
I’ve submitted my GSoC 2026 proposal for the AStats project and would really appreciate your feedback whenever you have time.
While preparing the proposal, I focused specifically on:
-
Designing structured, recipe-driven statistical workflows
-
Ensuring tool-grounded reliability for statistical correctness
-
Incorporating human-in-the-loop interaction
-
Exploring open-weight model usage and workflow predictability
I also built a small exploratory prototype to better understand the orchestration and design challenges, but I’ve treated it only as a baseline and expect to refine the approach based on your guidance.
If possible, I would especially value your thoughts on:
-
Whether the proposed workflow structure aligns with your expectations
-
The evaluation strategy for reliability and usability
-
The balance between automation and practitioner control
Thank you for your time I’m really excited about this project and would love to iterate based on your feedback.
Best regards,
Abhishek Buddiga
1 Like
I’ve finished a detailed draft for the AStats proposal. I focused on having reliability and editable structure. I cannot send a DM so if you could send me one first to unlock it or request access on the following link (you should also have it in your inbox)
1 Like
Please send me a link via DM.