TaskPilot AI

Inspiration

We wanted to build something that actually uses AI in a practical way, instead of just plugging in a model for every task. Most AI apps feel generic, but TaskPilot focuses on picking the right model for each feature to get the best results.

What it does

TaskPilot has three main features:

Chat – Real-time multi-turn chat with selectable Llama models and full Markdown rendering.
Translate – Text translation using Llama 3.3 70B, plus image translation with Llama 4 Scout’s vision capabilities to extract and translate text from photos.
Math – Upload a photo of a math problem; Llama 4 Scout extracts the problem, and Llama 3.3 70B solves it step by step, rendered neatly with KaTeX.

All features use only free-tier APIs, no paid services.

How we built it

Backend: Flask, deployed on Vercel as a Python serverless function.
Inference: Runs through Groq API for fast responses.
Model assignments:
- Llama 4 Scout handles vision tasks (image text extraction for translate and math).
- Llama 3.3 70B handles reasoning-heavy tasks (translation and math solving).
- Chat tab: Users can pick any model they want.
Frontend: Plain HTML, CSS, and JavaScript. Supports drag-and-drop images and real-time streaming for chat.

Challenges we ran into

Extracting math problems accurately without extra commentary was tricky.
Getting Llama 3.3 70B to produce clear, readable explanations with LaTeX formatting required fine-tuning.
Ensuring long outputs expanded naturally in result cards without clipping.

Accomplishments that we're proud of

Two-step math pipeline works smoothly end-to-end.
Clean and minimal interface despite using only vanilla JavaScript.
Publishing this app as a public website

What we learned

Choosing the right model for each task is more effective than using the largest model everywhere.
Llama 4 Scout is excellent at vision extraction; Llama 3.3 70B handles structured reasoning well.
Splitting tasks across models improves output quality significantly.