This is a submission for the GitHub Finish-Up-A-Thon Challenge.
🛡️ What I Built #
TruthGuard AI is an enterprise-grade fake news detection and fact-checking tool that combines real-time web search with advanced LLM analysis. Users can paste any claim, news excerpt, or social media post, and the tool:
- Searches the live web using Tavily's search API
- Analyzes evidence using Groq's Llama 3.3-70B model
- Provides a structured verdict with veracity score (0-100%)
- Identifies logical fallacies and emotional manipulation
- Offers a follow-up chat to interrogate the evidence
The application is built with Streamlit for the frontend, LangChain for LLM orchestration, Tavily for live web search, and Groq's Llama 3.3-70B for fast, high-quality inference.
This project started as a hackathon prototype where I had the core verification logic working, but it lacked persistence, testing, proper environment configuration, and a polished user experience. It was functional but incomplete — perfect for this challenge.
Demo #
🔗 GitHub Repository: https://github.com/Awaisranahmad/AI-Based-Content-Verification-Tool
GitHub Copilot in Action:
📸 Before vs After — The Comeback Story #
Where It Was (The "Before")
The original version of TruthGuard AI was a working prototype that proved the concept but had significant gaps:
| Issue | Impact |
|---|---|
| ❌ No history persistence | Results disappeared on page refresh |
| ❌ No .env support | API keys only worked via Streamlit secrets; local development required manual editing |
| ❌ No unit tests | Zero test coverage — fragile to API changes |
| ❌ No delete or load functionality | History items stuck forever |
| ❌ No "New Chat" option | No way to reset session |
❌ Deprecated st.experimental_rerun() |
|
| Caused AttributeError in newer Streamlit versions |
The code worked for me, but nobody else could use it reliably. There was no proper database setup, no environment configuration, and no way to revisit past verifications.
What Changed (The "After")
I transformed TruthGuard AI from a prototype into a production-ready application with GitHub Copilot writing most of the new infrastructure.
| Feature | Before | After |
|---|---|---|
| History | ❌ None | ✅ SQLite database with timestamps |
| Load old claims | ❌ Not possible | ✅ One-click load with full report |
| Delete history | ❌ Not possible | ✅ Delete button per item |
| New chat | ❌ No reset | ✅ Reset all session state |
| Environment | ❌ Hardcoded/st.secrets only | ✅ .env + st.secrets fallback |
| Unit tests | ❌ 0 tests | ✅ 3 passing tests |
| API calls on load | ❌ N/A | ✅ Load uses saved report (no re-verification) |
| Deprecated code | ❌ experimental_rerun |
|
✅ rerun |
||
New Files Added
During the revival, I added 6 new files to the repository:
history_manager.py
— SQLite database operations (init_db, save_history, get_all_history, delete_history) -
tests/test_history_manager.py
— Unit tests with 3 passing tests -
.env.example
— Environment variable template -
.gitignore
— Python-appropriate ignores (.env,pycache, *.db) - Updated
requirements.txt
— Added pytest and python-dotenv - Updated
app.py
— Integrated history, .env support, and new UI buttons
Verdict Summary Extraction Logic
One of the more interesting additions was the extraction of structured data from the LLM's markdown output. The extract_verdict_summary()
and extract_veracity_score()
functions now cleanly parse the Groq Llama 3.3 response into database-friendly fields — something Copilot helped generate in a single prompt.
My Experience with GitHub Copilot #
This was my first time using GitHub Copilot for a serious revival project, and it completely changed how I approach incomplete work. Here's exactly how Copilot helped me:
1. Copilot Built the Entire history_manager.py
Module
My prompt:
"Create a Python module with SQLite database functions to store fact-check history. Columns: id, timestamp, claim, verdict_summary, veracity_score, full_report_text. Include functions: init_db(), save_history(), get_all_history(), delete_history()."
Copilot generated: All 6 database functions with proper error handling, SQL injection prevention (using parameterized queries), and connection management. I didn't write a single line of SQL manually.
2. Copilot Added .env Support with Fallbacks
My prompt:
"Add code to load .env file using python-dotenv, then fallback to st.secrets and os.getenv for API keys."
Copilot generated: The complete environment chain, handling local development and Streamlit Cloud deployment seamlessly.
3. Copilot Wrote All 3 Unit Tests
My prompt:
"Write pytest unit tests for history_manager.py – test init_db, save_history, get_all_history, delete_history. Use temporary file for database."
Copilot generated: Complete test file with fixtures and assertions. All 3 tests passed on first run.
4. Copilot Added the "New Chat" and "Delete" Buttons
My prompt:
"In the sidebar history section, next to each history item, add a small delete button. Also add a 'New Chat' button that clears the session and reruns."
Copilot generated: The full Streamlit UI code with proper column layouts and session state management.
5. Copilot Fixed the Deprecated experimental_rerun
I ran into the AttributeError: module 'streamlit' has no attribute 'experimental_rerun'
error. Instead of searching Stack Overflow, I asked Copilot:
"What replaces experimental_rerun in newer Streamlit versions?"
Copilot answered: st.rerun()
and replaced all occurrences automatically.
Key Copilot Patterns I Learned
| Pattern | How I Used It |
|---|---|
| Comment-driven prompts | |
| Writing detailed comments before implementation guides Copilot to generate exactly what's needed | |
| Function signature first | |
Typing def save_history( and letting Copilot infer the schema |
|
| Iterative refinement | |
| Accepting suggestions, then asking "Add error handling" or "Use parameterized queries" to improve them | |
| Test generation | |
| Asking Copilot to write tests before implementing the function (test-driven development with AI) |
The biggest surprise:Copilot didn't just autocomplete — it understood the project structure (Streamlit, SQLite, pytest) and generated context-aware code that integrated seamlessly with my existing logic.
🛠️ Technical Stack #
| Component | Technology |
|---|---|
| Frontend | Streamlit (Python) |
| LLM | Groq API (Llama 3.3-70B-Versatile) |
| Search | Tavily Search API |
| Orchestration | LangChain |
| Database | SQLite |
| Environment | python-dotenv |
| Testing | pytest |
| Dev Assistant | GitHub Copilot |
🔍 Judging Criteria Reflection #
| Criterion | How TruthGuard AI Addresses It |
|---|---|
| Use of underlying technology | |
| Combines Groq's LPU (sub-second inference), Tavily's real-time search, and LangChain for structured prompting | |
| Usability & UX | |
| Glassmorphism UI, progress indicators, chat-based follow-ups, one-click report downloads, persistent history | |
| Originality & Creativity | |
| LLM-powered fact-checking with structured verdicts (Veracity Score + Final Verdict + Bias detection) is not a basic chatbot wrapper — it's a purpose-built misinformation tool | |
| Completion Arc | |
| Clear "before" vs "after" transformation: from prototype to database-backed, tested, environment-aware application |
📦 How to Run Locally #
bash
git clone https://github.com/Awaisranahmad/AI-Based-Content-Verification-Tool.git
cd AI-Based-Content-Verification-Tool
python -m venv .venv
source .venv/bin/activate # On Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env
streamlit run app.py