Inspiration

As university students taking classes in an international environment, we often struggle with unclear lectures, fast-paced explanations, and complex topics delivered in varying levels of English. Our notes would be messy, incomplete, or factually uncertain and verifying every detail became a time-consuming task we kept putting off.

We asked ourselves:

“What if notes could correct themselves?”

That question became the foundation of TruthLens which is a tool built for students like us who want to learn faster, more accurately, and with more confidence.

What it does

TruthLens is an AI-powered study workspace that analyzes your notes sentence-by-sentence, flags inaccuracies, explains why they’re incorrect, and proposes corrected versions backed with credible sources.

With one click, you can apply these corrections directly into your document, creating a polished, fact-checked version of your notes.

In short:
TruthLens turns raw notes into reliable knowledge.

How we built it

Frontend:
Built with Next.js 16, React 19, TypeScript, Tailwind, and shadcn/ui for a modern, clean interface.

Backend:
Django REST Framework powers user auth, document storage, sentence parsing, and correction workflows.

AI Engine:
We integrated Ollama to run a local LLM with a custom fact-checking prompt.
The model:

Evaluates each sentence
Labels it as true/false
Proposes a correction
Explains its reasoning
Provides source links

Data Storage:
PostgreSQL stores users, documents, sentences, and corrections.

Smart Correction System:
When a user clicks “Apply Correction,” the backend rewrites only the relevant sentence and regenerates updated sentence boundaries automatically.

Dev Environment:
Fully containerized with Docker for easy setup and cross-platform reliability.

Challenges we ran into

Sentence Matching: Ensuring the AI’s returned sentence exactly matched the extracted database version required multiple iterations.
Applying Partial Corrections: Replacing only one sentence inside a larger document while keeping indexes accurate ; it was unexpectedly complex.
LLM Output Parsing: LLMs often return malformed JSON, so we built sanitization and validation layers.
Real-Time UX: Coordinating loading states, highlights, and correction previews required careful UI design.
Local LLM Performance: Running fact-checking models locally with Ollama needed optimization to keep analysis smooth. It also needed RAM, so only one of us could test it.

Accomplishments we’re proud of

Built a fully working AI fact-checking system , not only a prototype.
Integrated a local LLM with structured, predictable outputs.
Designed a polished UI showing flagged sentences, explanations, sources, and instant corrections.
Created a robust correction workflow with proper sentence rewriting and regeneration.
Wrote strong GitBook documentation.
Overcame heavy backend/AI parsing bugs and shipped something reliable.

What we learned

How to build AI-powered features that feel useful, not gimmicky.
How to integrate LLMs safely and predictably. As well as for free, didn't spend a cent on AI.
The importance of clean prompt engineering and output validation.
How to handle document segmentation, dynamic rewriting, and sentence index recalculation.
How to coordinate frontend–backend–AI workflows effectively.
That small UX touches dramatically improve the user experience of AI features.