pplonski86

Thank you! I have also CSV dataset about tech layoffs, it would be super interesting to connect both in one analysis https://github.com/mljar/mercury-examples/tree/main/layoffs

I was able to fit only small models Qwen3.5-4B in RTX3070 which is not very useful for Python and SQL generation thought. When I wan to test larger open LLM models I often just use cloud resources.

Yesterday I installed llama.cpp to test it with local AI Data Analyst that I'm building. I was also testing other open LLM providers: Ollama, Jan, vLLM, LM Studio. I had older NVIDIA card (RTX 3070) and llama.cpp instalation was smooth, contrary to vLLM which required me to reinstall CUDA drivers because by default it installed the latest one. I'm curious if there is a speed difference between the same open LLM model served with different runners.

is it terminal on steroids some kind of? so you can manage mutiple coding agents? how many coding agents you can manage in parallel that it is still comfortable to work and code changes are meaningful

I love such puzzles! I used to use a lot ternary operators in C++ but one day friend of mine told me that I shouldn't nest ternary operators too much because code is too complicated to read - he understands code perfectly, he was just worried about younger programmers. Since then I started to use longer versions of code instead of smart shortcuts - to improve readability of code.

before I ask AI to write anything, I prepare a plan, I was very positively surprised when noticed Plan mode in Codex recently. It make me feel that maybe others doing the same and that's why they added it. Anyway, I start with plan, then ask AI to do just one step.

If coding a new feature, I do one step and check the code, doing git diff, reading changes, or just asking Codex, to show me changes.

If writing an article, I ask for only one paragraph. I read paragraph and if it is ok, I accept it, if it doesn't show off my thoughts I work on one paragraph.

If doing data analysis with AI, I do one step of analysis and ask AI to display intermediate results so I can see if all is going in good direction and there are no hallucinations, additionally I have follow-up prompts for AI to do results verification. If all looks good, then I continue to the next step.

I don't like situation when I ask AI to do all code changes, or all article, or all data analysis in one pass with one prompt. It is simply impossible to check if AI is correct and results are not satisfactory. You can easily see this when asking AI to write a deep article with one prompt - you clearly see that it doesn't reflect your thoughts.

Maybe step-by-step is the approach to use AI and not feel dumber.

Python outputs is also very versatile. You can use Python to build command line script, web application, desktop app with GUI, notebook with data analysis, or Python package and share with others. It is many ways how Python code can be used by final user.

Im working on AI data analyst - MLJAR Studio. It is conversational UI with AI agent which uses Python to provide data insights. It is available as desktop application https://mljar.com

I'm working on an AI Data Analyst in MLJAR Studio and found this small but interesting bug while testing a medical data use case. The AI-generated Pandas code looked correct and executed without errors, but the dataframe was misaligned. The first patient had 148 pregnancies because glucose values were shifted into the Pregnancies column. The interesting part for me was that the bug was caught only because both the displayed dataframe and an extra LLM output checking step were reviewed.

Thank you! I will check them out. It is worth to mention that MLJAR Studio is a desktop application, which is easy to install. It is running locally, and support local LLMs so all data stay safe.

User is not touching notebook at all, user just ask questions in natural language, and AI is using Python to compute answer, the ipynb notebook format is used to save the conversation.

Human in the loop in data analysis is really challenging task. We provide Python code for inspection, so user can check details how results were produced. Additionally, we run AI on results - user need to check the outputs and AI provided insights.

Thanks for sharing! MLJAR Studio was created for people with domain knowledge but not much technical expertise. For them, setting up a Python environment, installing required packages, configuring Jupyter Lab, the MCP server, and Claude Code might be technically demanding.

MLJAR Studio is a desktop application available for Windows, MacOS, and Linux. MLJAR Studio creates a Python environment for the user and installs all required packages. The user can focus on data rather than fighting technical challenges.

The goal of MLJAR Studio is to make it easy to analyze data for people with large domain knowledge but lack of programming skills. We do not focus on notebooks. Python notebook for us is compute and store layer. Our main interface is chat with AI data analyst. The conversation can be opened as classic notebook, but the main UI is simple chat.

Python notebooks are not reproducible when used by humans. When notebook format is used to store conversations for AI data analysis, it preserves the chat history and is ideal for reproducibility.

I like Ollama Cloud service (I'm paid pro user), because it let me test several open source LLMs very fast - I dont need to download anything locally, just change the model name in the API. If I like the model then I can download it and run locally with sensitive data. I also like their CLI, because it is simple to use.

The fact that they are trying to make money is normal - they are a company. They need to pay the bills.

I agree that they should improve communication, but I assume it is still small company with a lot of different requests, and some things might be overlooked.

Overall I like the software and services they provide.

We built a benchmark to evaluate LLMs on real data analysis workflows. Instead of single prompts, each task is a sequence of prompts (steps). It is similar to how a human data analyst works in practice. Each run is saved as full python notebook, including prompts, code and outputs. We evaluated runs across task completion, code correctess, output quality, reasoning and reliability. Each workflow is execuuted multiple times and scored automatically.

Modern LLMs perform very well on individual steps. The benchmark currently inludes 23 workflows from different data analysis tasks (EDA, ML, NLP, statistics ...). The top-3 models across the 23 workflows, gpt-oss:120b scored 9.87/10, followed by gpt-5.4 at 9.65/10, glm-5.1 at 9.48/10. Which is very high in my opinion. The results show that modern LLMs perform very well on data analysis tasks. All feedback is welcome! I uploaded all notebooks for each model https://github.com/pplonski/ai-for-data-analysis

I thought it was open source project on github? https://github.com/anthropics/claude-code no?

Can someone explain ELI5 how it does work? and how many data points it can read?

Do we need rockets to put satelittes to the space? Cant it be done with baloons? https://www.youtube.com/watch?v=NFieAD5Gpms

It is not that easy to build such app from scratch ... it all requires a lot of work, even with AI help. I think the most important is to provide easy to use UI first, and if speed or some missing features will be blockers for further innovation step then maybe native app will be at some point created.

I have dual boot on decent laptop, doing nothing, on windows fan is always on, computing something? On Linux it is just silent

Thank you! With WASM I can’t use all pypi packages and can’t connect to database, that’s why I’m looking for python based solution

Thank you! Exactly what I was looking for

What if AI starts to have sense of craft? we just miss the verify and critique models, that will tell other models what looks good

thank you for sharing, is there a new container for each code run, or it stays the same for whole conversation?

There are so many models, is there any website with list of all of them and comparison of performance on different tasks?

Maybe point 9, trust but verify, should be extended to AI coworkers as well. I would love to have tools to verify AI code by quantity.

Are you aware of any lightweight sandboxes for Python? not browser based

very nice comparison! I'd like to see on what examples OCR engines fail