The latest articles on DEV Community by 唐远韬 (@yuantao). https://dev.to/yuantao 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%2F4178212%2F6c9b969e-afb8-497f-a32a-3bd1d1261ed0.png https://dev.to/yuantao en 唐远韬 Mon, 12 Oct 2026 04:01:50 +0000 https://dev.to/yuantao/i-built-intentlens-to-make-metric-definitions-visible-before-generating-charts-56i6 https://dev.to/yuantao/i-built-intentlens-to-make-metric-definitions-visible-before-generating-charts-56i6 <p>A chart can look right while answering the wrong question.</p> <p>For example, imagine two groups:</p> <ul> <li>Group A: 10 conversions from 100 visits → 10%.</li> <li>Group B: 5 conversions from 10 visits → 50%.</li> </ul> <p>The overall conversion rate is <strong>15 / 110 ≈ 13.64%</strong>.<br> The average of the two group rates is <strong>30%</strong>.</p> <p>Both calculations are valid, but they answer different questions. I wanted that choice to be visible before a chart is generated.</p> <h2> What I’m building </h2> <p><strong>IntentLens</strong> is a local Excel/CSV visualization workbench.</p> <p>The workflow is:</p> <p><strong>Import data → Describe the question → Review a plan → Compute → Refine → Export</strong></p> <p>Before execution, you can inspect grouping, aggregation, ratio definitions and missing-value handling. The full application also supports chart revisions and reproducible exports.</p> <p>Models propose declarative analysis plans rather than arbitrary Python or JavaScript to execute. The application validates those plans and computes results from a data snapshot.</p> <h2> Try the English interactive sample </h2> <p><a href="https://pyslowpoke.github.io/IntentLens/?lang=en" rel="noopener noreferrer">Open the sample</a></p> <p>No installation, signup or API key is required.</p> <p>Try one small task:</p> <ol> <li>Select the conversion-rate question.</li> <li>Compare overall conversion with average monthly conversion.</li> <li>Generate a chart.</li> <li>Change its title or ordering.</li> <li>Download a result.</li> </ol> <p><strong>The public sample uses fixed synthetic data and deterministic rules. It does not call AI or accept file uploads.</strong></p> <p>The two-group example above explains the concept; it is separate from the sample’s 12-record dataset.</p> <h2> What the full local application adds </h2> <p>The full workbench supports your own Excel/CSV files, conversations with a configured model, multiple chart engines and revision history.</p> <p>I also recorded an English walkthrough of the actual local workflow: uploading a synthetic CSV, entering a question, receiving a plan from DeepSeek and generating a chart. The computed regional totals were checked against the input data.</p> <p>The project is still a development version intended for trusted local, single-user use. Model-assisted planning requires an API key and sends field metadata and statistics to the configured provider.</p> <p><a href="https://github.com/Pyslowpoke/IntentLens" rel="noopener noreferrer">Source code and installation instructions</a></p> <h2> The feedback I need </h2> <p>I’m trying to understand whether the confirmation step earns its place in a real reporting workflow.</p> <p>If you regularly work with spreadsheets:</p> <ul> <li>For your last report, would reviewing the metric definition have prevented a mistake or added friction?</li> <li>Where did you get stuck in the sample?</li> <li>What would you use instead, and what does that workflow do better?</li> </ul> <p>I’m also looking for five people with recurring reporting tasks to try the full application using public or sanitized data.</p> <p>Comments and concrete criticism are welcome. No star required.</p> ai python webdev productivity