The latest articles on DEV Community by Arun Mahajan (@arun2232). https://dev.to/arun2232 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%2F4158829%2F0751e5f7-ed96-4061-a1e6-341e979adef3.jpg https://dev.to/arun2232 en Arun Mahajan Sat, 03 Oct 2026 12:55:47 +0000 https://dev.to/arun2232/udhaar-a-private-voice-khata-for-a-kirana-shopkeeper-5431 https://dev.to/arun2232/udhaar-a-private-voice-khata-for-a-kirana-shopkeeper-5431 <h1> Udhaar: a private voice khata for a kirana shopkeeper </h1> <p><em>Submission for the <a href="https://dev.to/challenges/hacktoberfest-weekend-2026-10-01">Hacktoberfest Weekend Challenge: Build for a Friend</a>.</em> </p> <h2> What I built </h2> <p>I built <strong>Udhaar</strong> for my friend <strong>Purushottam</strong>, who runs a neighbourhood kirana shop in <strong>Pune</strong>. His recurring problem was recording small customer debts quickly while serving the next person. Paper notes get lost, and a cloud app is a poor fit for sensitive customer names and balances.</p> <p>Udhaar lets the shopkeeper speak or type a Hindi, Marathi, or Hinglish entry such as “Ramesh ji ko 340 ka samaan udhaar”. It proposes a structured entry, asks the shopkeeper to confirm it, and then updates a simple khata.</p> <p>It also shows balances, a monthly summary, and polite WhatsApp reminder links. The links are always manually opened and sent; the app never messages a customer automatically.</p> <h2> Demo </h2> <ul> <li>Demo video: <a href="https://youtu.be/8X4TxyVbnI8" rel="noopener noreferrer">https://youtu.be/8X4TxyVbnI8</a> </li> <li>Repository: <a href="https://github.com/aun009/udhaar" rel="noopener noreferrer">https://github.com/aun009/udhaar</a> </li> <li>Local demo: <code>UDHAAR_MOCK=1 uvicorn app.main:app --host 0.0.0.0 --port 8000</code>, then open <code>http://127.0.0.1:8000</code>.</li> </ul> <p>The mock mode runs without a GPU, microphone, or network model service. The production path uses local Whisper and Ollama instead.</p> <h2> How I built it </h2> <p>The system is a small FastAPI application with SQLite storage and a single mobile-friendly HTML page:</p> <ol> <li>Audio is transcribed locally with <code>faster-whisper</code> when voice mode is enabled.</li> <li>An open-weight Qwen instruct model served by Ollama converts the transcript into a strict JSON label.</li> <li>The UI displays the proposed customer, amount, type, and note for confirmation.</li> <li>Confirmed entries are stored locally in SQLite and used for balances and summaries.</li> <li>A deterministic Hindi/Marathi reminder template produces a WhatsApp link for manual sending.</li> </ol> <p>The repository also includes a synthetic Hindi/Marathi/Hinglish dataset generator, an evaluation runner, and a heuristic fallback so the core demo remains usable when Ollama is unavailable.</p> <h2> Why open innovation mattered </h2> <p>Open models made the important parts adaptable instead of hiding them behind a remote API. Customer names and dues can remain on the shop laptop, amount-word examples can be added to the training set, and the parser model can be swapped or fine-tuned. Once the local runtime is installed, there is no per-entry API bill and the workflow can continue when the internet is unreliable.</p> <p>That control has a cost: local model setup is more involved, and zero-shot parsing is less robust than a strong hosted model on messy speech. For this friend and this data, keeping the ledger private and editable was the better trade.</p> <h2> Results and limitations </h2> <p>The reproducible heuristic baseline currently scores:</p> <div class="table-wrapper-paragraph"><table> <thead> <tr> <th>Split</th> <th>Examples</th> <th>Exact JSON</th> <th>Customer</th> <th>Amount</th> <th>Type</th> <th>Valid JSON</th> </tr> </thead> <tbody> <tr> <td>Synthetic test</td> <td>400</td> <td>16.00%</td> <td>47.50%</td> <td>48.00%</td> <td>94.00%</td> <td>100.00%</td> </tr> <tr> <td>Unseen templates</td> <td>120</td> <td>10.00%</td> <td>25.83%</td> <td>53.33%</td> <td>93.33%</td> <td>100.00%</td> </tr> </tbody> </table></div> <p>The second row should be read as 120 examples, 10.00% exact JSON, 25.83% customer, 53.33% amount, 93.33% type, and 100.00% valid JSON. These are baseline numbers, not a claim that the parser is solved. Ollama, fine-tuning, and real-audio rows are intentionally omitted until they are run locally.</p> <h2> Open source and code </h2> <ul> <li>Source: <a href="https://github.com/aun009/udhaar" rel="noopener noreferrer">https://github.com/aun009/udhaar</a> </li> <li>Evaluation details: <code>eval/RESULTS.md</code> </li> <li>Training notes: <code>train/README.md</code> </li> </ul> <h2> My agent session </h2> <p><strong>[Optional: add the DevRelay agent-session link or remove this section.]</strong></p> <h2> Partner categories </h2> <p><strong>[List only categories that actually apply, or remove this section.]</strong></p> devchallenge weekendchallenge hf26challenge