🚀 Built something cool - an AI-powered Text-to-SQL system Ask database questions in plain English and get instant SQL results—no complex queries, no API costs, and full data privacy. ✨ What it does: • Converts natural language → SQL using local AI (Ollama) • Lets you upload your own data (CSV / Excel / JSON / SQL) • Switch databases instantly with live schema updates • Clean UI with dark/light mode • Secure, SELECT-only query execution 🧠 Why it’s different: Most demos use fixed datasets. This one understands your data and generates accurate queries in real time. 🛠 Built with: React, Tailwind CSS, FastAPI, SQLite, Ollama 🔗 Open-source | Free | Privacy-first Check out the demo and code on GitHub 👇 https://lnkd.in/gS6nPjNs #AI #GenAI #FullStack #SQL #WebDevelopment #OpenSource #React #Python

Last year I built a Text-to-SQL agent, using Langchain. In order to make it work, I had to add a lot of python helper functions. Five months later I rebuilt the same system using -Agent Skills (introduced by Anthropic) -MCP BigQuery connector, right from the chat in Claude -And queried directly from Claude interface. My entire codebase just got replaced by a simple markdown file describing the data dictionary and constraints. The connector helped to link with the database. The skill.md file held the contextual knowledge. Claude provided the reasoning power. Key lesson: • Not every problem needs an agent. Many just need the right capability. • Reliability comes more from clear context than clever prompting. • With Agent Skills, domain experts -not just developers, can build useful AI systems. Link to my article in comments below.

📅 Day 100 of 100 Days of Data Science 🎥🤖 Topic: The Autonomous Data Scientist — AI Agent for CSV-to-Insights (Gemini + E2B + Streamlit) 🚀 What I built For the final day of my challenge, I built a Streamlit app that turns any uploaded CSV into an interactive “autonomous analyst.” Instead of manually writing notebooks, you ask questions in plain English and the agent generates Python code, executes it in a sandbox, fixes errors, and returns results with charts—end to end. ✅ Core features - CSV upload → instant workspace: works with any tabular dataset you provide - Agentic analysis loop (Gemini): converts natural language requests into executable Python (EDA, visualization, modeling) - Safe code execution (E2B sandbox): runs generated code in an isolated environment - Self-correction: if code fails, the agent reads the error, fixes the script, and retries automatically - Transparent outputs: shows executed code + logs for reproducibility and trust - Visualization-first: generates relevant Matplotlib/Seaborn charts and renders them directly in the app 🧠 Why this matters Most analytics tools still require you to switch between asking questions and writing code. This project explores an agentic workflow where an LLM can: 1) reason about the dataset, 2) write analysis code, 3) execute it safely, and 4) debug itself—making data exploration faster and more accessible while keeping the process auditable. 🧰 Tech stack Streamlit, Pandas, Matplotlib, Seaborn, Google Gemini, LangChain, E2B Code Interpreter (sandboxed execution), Python (Colab deployment) 🤖 GitHub Github Link:-https://lnkd.in/ghHb3SsH #DataScience #AI #LLM #AgenticAI #Python #Streamlit #Pandas #DataAnalytics #AutoML #MLOps #100DaysOfDataScience #PortfolioProject

What does it mean to be "good at data"? The industry is currently obsessed with tools—SQL, Python, AI—but we’re starting to overlook the actual thinking that makes those tools useful. I’d argue those tools are just the baseline; the real skill is navigating a methodology that is far from linear. In data science/analytics, we don't follow a straight path. The work is defined by a unique phenomenon, for example: - Multiple Paths: There are often 100 different ways to approach the same dataset. - Variable Truths: There can be 100 "correct" answers depending on the business lens or timeframe applied. - Pattern Recognition: Sometimes, 100 seemingly unrelated problems share a single, underlying root cause. Navigating these 100 possibilities isn't a technical skill, 𝐢𝐭’𝐬 𝐚 𝐜𝐫𝐞𝐚𝐭𝐢𝐯𝐞 𝐨𝐧𝐞. Most people think data work is rigid (gabage in, gabage out), but the real secret sauce is 𝘴𝘵𝘳𝘶𝘤𝘵𝘶𝘳𝘦𝘥 𝘴𝘬𝘦𝘱𝘵𝘪𝘤𝘪𝘴𝘮. It’s about playing the Devil’s Advocate against your own findings and having the curiosity to question the question before you even touch a line of code. Sure, we all need to know how to program. That part is a given. But the real value comes from 𝐭𝐡𝐢𝐧𝐤𝐢𝐧𝐠 𝐨𝐮𝐭𝐬𝐢𝐝𝐞 𝐭𝐡𝐞 𝐛𝐨𝐱, this is the difference between someone who can run a script and someone who can actually invent a solution for a problem that didn't exist yesterday. Next time you approach a data problem, don't just reach for your favorite tool. Start by embracing 𝐭𝐡𝐞 100-𝐰𝐚𝐲 𝐟𝐫𝐚𝐦𝐞𝐰𝐨𝐫𝐤 . Be skeptical, ask the wrong questions, and allow yourself to be a bit more creative. #DataAnalytics #ProblemSolving #DataMethodology

𝗖𝗹𝗮𝘂𝗱𝗲 𝗖𝗼𝗱𝗲 𝗷𝘂𝘀𝘁 𝗰𝗵𝗮𝗻𝗴𝗲𝗱 𝗵𝗼𝘄 𝗺𝘆 𝗮𝗻𝗮𝗹𝘆𝘁𝗶𝗰𝘀 𝗯𝗿𝗮𝗶𝗻 𝘄𝗼𝗿𝗸𝘀. I downloaded Claude Code into VS Code and tried something simple. Plain English. No Python. No SQL. And… it worked! I loaded a CSV and instantly understood the structure. Cleaned messy columns. Handled missing values. Standardized formats. Performed keyword search. 𝗛𝗼𝘄 𝗶𝘁 𝘄𝗼𝗿𝗸𝘀 𝗶𝗻 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲: Inside VS Code, I just asked Claude Code something like: Load my sales data and show me: - Top 10 customers by revenue - Regional performance - Month-over-month growth That was it. Claude Code wrote the Python, ran it, and showed the results. I didn’t touch the code. I didn’t need to know the syntax. It felt less like “programming” and more like having a conversation with my data. That part surprised me. What surprised me more was the shift in mindset. I wasn’t thinking about syntax. I wasn’t debugging. I was thinking about the data itself. The time savings are obvious. But the bigger change is focus. Less time telling the computer how to do things. More time thinking about what the data is saying. I’m still early, but I’m excited to try this next: - Predictive analysis - Statistical exploration - Other ML style analysis, asked in plain language One more thing I really like: The data stays on my MacBook. It’s fast. It feels private. It feels… safe. I don’t think tools like this replace data analysts. I think they quietly remove the friction that never added value in the first place. 𝗖𝘂𝗿𝗶𝗼𝘂𝘀: What data analytics use cases are you using Claude Code for?

I worked on a common problem in any organization - to find insights in data stacked on csv or any table. While we have usecases such as text2sql to query structured data in database, most times tables we want answers from on daily life is in csv or excel. Introducing talk2table:   Intelligent Data Understanding:  When you upload a CSV, a sophisticated data scientist agent meticulously scans every column. It identifies ranges, missing values, categories, and truly "gets to know" your data – even incorporating your descriptions for deeper context. You can also test hypothesis on the go, generate charts on the go or fit any ML model on the go! Dynamic Querying & Insights:  Ask any question, and talk2table springs into action! It spins up a secure Python sandbox, loads your CSV as a DataFrame, and iteratively generates the precise Python commands needed to extract the answers you're looking for. Our orchestrator ensures we hit the mark, guiding the process with intelligent next steps until your query is resolved. Most importantly, you can see it working live - keeping all operations transperant to you. No more tedious scripting or waiting for data reports. With talk2table, you're empowered to uncover hidden trends, validate hypotheses, and make data-driven decisions faster than ever before. #AI #DataScience #MachineLearning #CSV #DataAnalytics #Innovation #talk2table

I feel claude code is partially broken for data science. Am I the only one? I've been using Claude Code extensively. For full-stack development they're incredible - I can scaffold apps, refactor codebases, write tests at 10x speed. But for data science? Its borderline useless. And I've spent weeks trying to figure out if it's me using it wrong. The core problem I feel agents are built for "build mode" (linear, requirements → implementation → done) but data science is "explore mode" (iterative, hypothesis → experiment → learn → pivot → repeat). What I keep hitting 1. The Eager Executor Problem - I want to think about my hypothesis. Claude wants to immediately generate a "complete analysis pipeline with visualizations." Bro, slow down, I don't even know what question I'm asking yet. 2. Context Decay - By interaction 5-6, my carefully established constraints ("only Q4 data", "target is churn not revenue") start getting ignored. It's like my CLAUDE.md never existed. 3. No Experiential Memory - "We tried random forest yesterday, it overfit." Claude next session: "Have you considered random forest?" There's no memory of what was tried. What I've tried: - Spec files in .claude/specs/ (helps, but I'm now the memory system) - Aggressive /compact usage (extends sessions, doesn't help across days) - Session summaries saved to files (lossy, manual, tedious) My conclusion: These agents fundamentally lack: - Experiential memory (what worked/failed) - Hypothesis tracking (what we're testing) - Constraint enforcement (what NOT to do) They have file system access and CLAUDE.md. That's it. Everything else is on you to maintain manually. I did try gsd and its pretty effective ngl but still as a Data Scientist I feel the way claude code or any other tool approcahes the problem is flawed. Am I missing something? How are other data scientists using these tools effectively? Any leads will help tbh, on another note I am thinking of building a meta prompting and context compression add on to claude code any one who has great ideas please DM! #claudecode #cursor #zerveai #gemini #geminicli #claude #anthropic

Everybody is writing about AI (LLMs) replacing software engineers. And I have to admit: with Claude Code you can easily create decent-looking web pages. But what about building something else? So I tried to build a simple data pipeline using Claude. Nothing fancy, just ingesting CSV files from a GCS bucket into a BigQuery table. The result? Something that technically works, but it is below junior data engineer level and not even close to production-ready. It had: - No clue about pagination - No clue about table partitioning - No clue about idempotency - No clue about async coding - No clue about anything that makes it performant and scalable But that’s no surprise. The model hasn’t seen enough of these real-world patterns to reliably reproduce them. It is not intelligent; it is still just a language model. So if you are doing more than “just” building websites, do not worry at all :) The real leverage is not typing code faster. It is knowing what to build, how to build it, and how to keep it running when things break. Your knowledge is what matters the most. Of course, LLMs help you move quicker no matter what, but you are still the one who gets it over the line. #dataengineer #python #ai #dataengineering #claude #claudecode #dataengineering

💡 "Show me sales by region" → Instant chart. No code. No queries. That's the power of AI-driven data analysis I just shipped. I built SmartAnalyst because I was tired of the same workflow: load data → write pandas → debug → visualize → repeat. What if we could skip straight to insights? Here's what it does: 🔹 Upload CSV → Ask questions in plain English 🔹 GPT-5 generates & executes pandas code automatically 🔹 Natural language converts to SQL queries 🔹 Auto-generates visualizations on demand 🔹 Complete data profiling in seconds Real example: ❌ Before: Write 15 lines of pandas, debug syntax, format output ✅ Now: "What's the average revenue by quarter?" → Done. The stack: FastAPI + React + OpenAI GPT-5 + Pandas + SQLite Why this matters: Data analysis shouldn't require memorizing syntax. The barrier between question and insight should be conversation, not code. This is democratizing data analysis - analysts move faster, non-technical teams get self-service insights, and everyone focuses on decisions instead of syntax. What I learned building this: → Designing secure AI-powered APIs → Prompt engineering for reliable code generation → Building sandboxed execution environments → Creating intuitive data experiences Version 1.0 is live. Already planning v2 with multi-file support, collaborative features, and advanced ML insights. Try it yourself: https://lnkd.in/eQJ26B8G What's the biggest friction point in your data workflow? Drop a comment - I'd love to hear your thoughts. #DataScience #AITools #Python #DataAnalytics #OpenAI

I’m currently building my foundation in Machine Learning. While the algorithms and models get all the glory, I’m spending my first week mastering the most critical foundation: data ingestion. In the real world, data is rarely "plug-and-play." It’s messy, fragmented, and lives in a dozen different formats. To build models that actually work, you first have to master the art of getting that data into the pipeline. Here is a breakdown of the three key areas I’ve mastered this week: 1. Advanced CSV & Flat File Handling I went beyond basic loading to handle complex, real-world scenarios: -Memory Management: Used chunksize to process massive datasets in manageable segments, preventing memory overflows. - Data Cleaning at Ingestion: Handled missing values using the na_values parameter and forced specific data types with dtype to ensure consistency. - Custom Parsing: Managed non-standard files by defining custom separators (like sep='\t' for TSV files) and manually assigning headers. -Feature Logic: Applied the converters parameter to transform data (like mapping long strings to abbreviations) as it’s being read. 2. JSON & SQL Integration Machine Learning requires data from diverse architectures: -Relational Databases: Utilized XAMPP to host a local MySQL environment, connecting Python to the database using mysql.connector. -Semi-Structured Data: Processed JSON files and fetched live HTTP data, such as real-time currency exchange rates, directly from URLs. 3. Building API Data Pipelines I built a functional pipeline to gather data from external web services: -The Request Cycle: Mastered the requests library to communicate with APIs and handle JSON responses. -Pagination Logic: Implemented loops to iterate through hundreds of API pages, ensuring the capture of complete datasets rather than just a single page. -Dataset Consolidation: Used pd.concat to merge dynamic data streams into a single, analysis-ready master file. All of this was developed and managed within Google Colab. #MachineLearning #Python #Pandas #LearningJourney