Forward-It

Forward it, Forget it.

Comment

Inspiration

In India, WhatsApp is where real life gets organised: school circulars, society notices, electricity bills and appointment SMSes all arrive there, buried between dozens of good-morning forwards. The people juggling all of it, often working parents, are exactly the people productivity apps leave behind. They won't install another app, type their tasks into it, or build a new habit around it.

TaskForge's theme, work should fit people, not force people to fit the tool, pointed to a simple idea: don't ask people to move their lives into a to-do app. Meet them where their work already is, and make the only action one they already know: forward.

What it does

Forward-It is a WhatsApp assistant. You forward it anything (a school circular, a bill SMS, a photo of a printed notice, a PDF or a voice note) and it:

finds every task inside: what to pay, bring, submit, wear or attend, with dates, times and amounts, and it quotes the words each task came from so you can trust it;

ignores the noise: good-morning forwards don't become tasks, and forwarding the same notice twice doesn't duplicate it;

reminds you at a useful time: usually 8 PM the evening before, while there's still time to act, plus a short summary every morning; speaks your language: English, हिंदी and Hinglish, including phrases like kal, parso and Friday tak; shares with family: one forward reaches both parents, reminders go to everyone, and when one person replies 👍 the other is told. No more "did you pay the fee?" calls. The website includes a live two-phone demo running the real bot, a family dashboard (Tasks, a 7-day timeline, and Impact numbers counted live), calendar export, and a guided demo that plays the whole story with captions.

How we built it

Bot core (Node.js + Express): one conversation engine shared by WhatsApp and the web demo. It handles commands (LIST, DONE, LATER, UNDO, SHARE, FAMILY), family "households" that merge two lists into one numbered list, and a scheduler for reminders and morning summaries.

AI (Groq): openai/gpt-oss-120b reads text and PDFs, qwen/qwen3.8-27b reads photos of notices, and Whisper (whisper-large-v3-turbo) transcribes voice notes. Every answer is forced into a strict JSON schema and validated again in code.

A deterministic safety net: language models are good at reading messy messages but unreliable at calendar arithmetic. The prompt includes the next 14 days, and code double-checks weekday names ("Friday tak" must land on a Friday) and day-of-month dates ("by the 15th" on the 24th means next month).

Offline fallback: a rule-based parser (chrono-node plus our own Hindi/Hinglish date words) keeps the bot answering without an AI key or when the AI is down. PDFs are turned into text locally with unpdf. Reminder timing: 8 PM the evening before if that's still ahead; otherwise 2 hours before (tasks with a time) or 7:30 AM that morning; otherwise within the hour.

WhatsApp (Twilio): a polling mode that works without a public URL, plus a signature-checked webhook mode.

Website: plain HTML, CSS and JavaScript modules with light and dark themes, live chat over Server-Sent Events, .ics calendar export and a private demo per visitor, deployable to Render with one render.yaml.

Testing: 59 automated tests, including fake Groq and Twilio servers.

Challenges we ran into

Indian dates and Hinglish: day-first dates (27-10-2026), "by the 15th", kal (which can mean tomorrow or yesterday), shaam 5 baje.

Models and calendars: the model sometimes placed "Friday" on a Thursday or chose a date that had already passed. A calendar in the prompt plus checks in code fixed both.

Free-tier limits: Groq's rate limits and an occasionally over-capacity vision model. The bot waits and retries once, then tells the user honestly instead of guessing.

WhatsApp access: Twilio's trial WhatsApp number only sends Twilio's own template messages. Forward-It receives and understands real WhatsApp messages, but it needs a full WhatsApp sender to reply with its own text. Polling mode removed the need for a public URL, and the web demo runs the identical bot code.

Sharing without confusion: merging two people's lists into one numbered list, and making a plain "DONE" mean the right task for whoever replies.

Accomplishments that we're proud of

One action, forward, is the whole interface: no app, no sign-up, no typing.

It reads text, photos, PDFs and voice notes, in three languages.

Family sharing turns a forwarded notice into a shared task with a clean handoff.

It's honest by design: every task shows its source quote, worked-out dates are marked, and it never invents tasks.

A guided demo anyone can watch on the website, with 59 tests behind it.

What we learned

Designing for people who won't change their habits means removing steps, not adding features.

Language models read messy, multilingual messages remarkably well, but the best results came from pairing them with small deterministic checks.

Trust matters more than cleverness: showing the original words next to each task makes AI output believable.

Platform rules (WhatsApp templates, 24-hour windows, trial limits) shape a product as much as the code does.

What's next for Forward-It

Approved WhatsApp templates, so reminders can arrive any time, not only within 24 hours of the last message.

Recurring tasks ("every month on the 5th") and assigning a task to one family member.

A weekly "how did we do" summary on WhatsApp, built from the Impact numbers.

Testing with real families and measuring missed deadlines before and after.

Built With

Share this project:

Updates

Submission history