Project TruthLens

Inspiration

In India, WhatsApp forwards are more trusted than newspapers for millions of people. Unfortunately, this has fueled mob violence, financial scams, vaccine hesitancy, and political unrest. Most fact-check tools are English-only, web-based, and inaccessible to rural citizens. We wanted to build something that is as easy as sending a forward but can actually stop misinformation before it spreads.

What it does

Citizens can send a suspicious forward (text, image, audio, or video) to our agent via WhatsApp/SMS/IVR. The agent: Understands the message in local language (via Granite multilingual models). Cross-checks it against trusted datasets (Kaggle fake news, government APIs, fact-check sites, credible news). Responds in local language with a verdict: ✅ True, ⚠️ Misleading, or ❌ Fake. Generates a counter-message/poster the citizen can easily forward back to their groups.

How we built it

Granite LLMs → Multilingual text & image understanding. Kaggle datasets → Fake News Detection Dataset, COVID misinformation, phishing links dataset, Hindi-English code-mix dataset. Agent Development Kit (ADK) → Built an agentic workflow: Input processing → Dataset/API cross-check → Classification → Counter-response → Shareable output. Data Preprocessing Toolkit → Cleaned, filtered, and benchmarked misinformation datasets. WhatsApp API + lightweight cloud backend → To make it accessible to feature phone users.

Challenges we ran into

Multilingual complexity → Fake news often comes in code-mixed Hindi-English or misspelled local scripts. Misinformation is fast-moving → Kaggle datasets cover past data; we had to combine it with real-time news APIs. User trust → Just saying “this is fake” isn’t enough; we needed clear counter-messages with sources. Scalability on low-end devices → Had to design lightweight, WhatsApp-first workflows.

Accomplishments that we're proud of

Built an end-to-end automated agent (not just an LLM demo). Successfully fact-checked multimodal forwards (text + image). Designed shareable counter-messages — making users part of the fact-check ecosystem. Made it multilingual (English + Hindi in MVP) — critical for India.

What we learned

Agentic workflows > Chatbots → The real power is in automating a full cycle (detect → reason → act → alert). Context matters → A claim isn’t just true/false; sometimes it’s misleading, and people need nuance. Accessibility is king → For public-good projects, WhatsApp/SMS is more impactful than fancy apps. Trust-building → AI outputs are only useful if people believe them → hence the need for verified sources.

What's next for Project TruthLens

Expand to more Indian languages (Tamil, Telugu, Bengali, Marathi). Add voice input/output for citizens who can’t type. Integrate with local fact-check NGOs + govt APIs for real-time updates. Create a public misinformation dashboard → track trending fakes region-wise. Partner with schools, cyber cells, and community centers for deployment.

Built With

• elasticsearch
• javascript
• langchain
• mongodb
• node.js
• python
• pytorch
• transformers