The latest articles on DEV Community by krishav js (@krishav_js_598bc485eaab5b). https://dev.to/krishav_js_598bc485eaab5b https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4170784%2F38a27f7d-e681-4090-b5b1-31d78799a9b5.png https://dev.to/krishav_js_598bc485eaab5b en krishav js Thu, 08 Oct 2026 09:51:10 +0000 https://dev.to/krishav_js_598bc485eaab5b/argue-with-my-data-1gl8 https://dev.to/krishav_js_598bc485eaab5b/argue-with-my-data-1gl8 <p>Most AI data analysts give you an answer. Ours tries to prove its own explanation wrong first.</p> <p>Inspiration</p> <p>AI data tools sound confident even when they're wrong. When revenue drops right after a big customer leaves, everyone blames that customer, even if they explain only a fraction of the drop. We wanted an analyst that questions its own explanations, shows its working, and admits when the data can't answer.</p> <p>What it does</p> <p>Upload a CSV or Excel file and ask "Why did revenue fall in March?". The app:</p> <p>calculates the real change in code<br> proposes competing explanations<br> fixes a pass/fail threshold for each before testing it<br> tests and verifies each one with deterministic analysis<br> labels each SUPPORTED, WEAKENED, REJECTED or UNTESTABLE, with an evidence trace</p> <p>On our demo data, the obvious suspect, a lost customer, explains only 8% of the drop (WEAKENED). The real driver is a product mix shift at 67% (SUPPORTED), and seasonality is honestly UNTESTABLE.</p> <p>How we built it<br> LangGraph orchestrates the investigation as a state machine.<br> Gemma 4 (via the Gemini API and LangChain) interprets the question, proposes hypotheses and acts as a critic. It never does the maths or changes a verdict.<br> pandas, NumPy and DuckDB compute every number, and Pydantic validates every model response.<br> Streamlit and Plotly show the verdicts, the evidence, and whether Gemma actually answered each step.<br> Challenges we ran into<br> A Gemma call that hung for minutes, which we fixed with timeouts and a safe fallback.<br> Keeping the AI from deciding verdicts after seeing the numbers.<br> Merging two versions of the project built in parallel.<br> Code that passed tests on one Python version and crashed on another.<br> Accomplishments that we're proud of</p> <p>It rejects the tempting explanation and finds the real one, with numbers anyone can verify. And it can say "I can't test this" instead of making something up.</p> <p>What we learned</p> <p>Plausible isn't the same as correct. Let the AI suggest ideas, let code do the maths, and always test on a clean machine.</p> <p>What's next</p> <p>A "challenge mode" that attacks its own conclusion, letting users argue their own theory, and support for longer data histories.</p> <p>Built with</p> <p>python, gemma, gemini-api, langgraph, langchain, pandas, numpy, duckdb, pydantic, streamlit, plotly</p> <p>Links<br> GitHub: <a href="https://github.com/whatsyourname909/hacktoberfest_-hack-day-coimbatore-x-init-club-and-idea-club" rel="noopener noreferrer">https://github.com/whatsyourname909/hacktoberfest_-hack-day-coimbatore-x-init-club-and-idea-club</a></p> ai webdev programming productivity