The latest articles on DEV Community by Krisha Shah (@krishashah03). https://dev.to/krishashah03 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%2F4161200%2F78500bb4-9b49-439e-869c-60939d5186b4.png https://dev.to/krishashah03 en Krisha Shah Sun, 04 Oct 2026 18:30:30 +0000 https://dev.to/krishashah03/fillmate-an-ai-add-in-for-my-hr-friend-4gej https://dev.to/krishashah03/fillmate-an-ai-add-in-for-my-hr-friend-4gej <p><em>This is a submission for the <a href="https://dev.tolink-to-challenge">Hacktoberfest Weekend Challenge: Build for a Friend</a></em></p> <h2> What I Built </h2> <p>A close friend of mine handles HR at a small company. Every new hire means opening the same Word salary template and copying details from emails by hand: name, designation, joining date, CTC, notice period. It's slow, and one wrong digit in a salary letter is a real problem.</p> <p>So I built <strong>AI Word Template Filler</strong>: a Word add-in where she pastes an email or a few lines of text, and it fills the template for her. Nothing is written silently. Every field shows up for review first, and anything missing is flagged instead of guessed.</p> <h2> Demo </h2> <p> <iframe src="https://www.youtube.com/embed/VrrP5rSS1sM" width="710" height="399"> </iframe> <br> [Short video: paste text → review panel → Apply to Word]</p> <h2> Code </h2> <p>[<a href="https://github.com/krishashah-03/superdocs-builds/tree/main/extensions/krishashah-03/word-fill-template" rel="noopener noreferrer">https://github.com/krishashah-03/superdocs-builds/tree/main/extensions/krishashah-03/word-fill-template</a>]</p> <h2> How I Built It </h2> <ul> <li> <strong>[Open-weight model, e.g. Llama 3.3 70B] via OpenRouter</strong> extracts fields from free text like emails and bullet points</li> <li> <strong>FastAPI</strong> backend normalizes dates, salary ("8 LPA" → 800000), and yes/no values, then detects gaps</li> <li> <strong>React + Office.js</strong> task pane shows a review screen; only approved fields are written into Word</li> <li>JSON/CSV input skips the AI entirely, so structured data stays deterministic</li> <li>63 pytest tests cover parsing, normalization, gap detection, and AI error handling</li> </ul> <h2> Why Open Models Matter Here </h2> <ul> <li> <strong>No lock-in:</strong> OpenRouter's OpenAI-compatible API let me test several open-weight models and pick the one best at structured extraction by changing one config line.</li> <li> <strong>A path to fully local:</strong> because the model is open-weight, the same setup can run on her own machine through Ollama later, so salary data never has to leave the office.</li> <li> <strong>Low cost:</strong> a small company can't justify an expensive API bill for filling forms; open models keep it nearly free.</li> </ul> <h2> What She Said </h2> <p>[I've sent her a demo, and she'll try it on the next new joiner's salary letter. I'll update this section with her honest verdict, good or bad]</p> <h2> What's Next </h2> <p>Offer letters and NDAs. The engine is schema-driven, so a new template only needs a new schema.json.</p> devchallenge weekendchallenge hf26challenge ai