• 

  Home page - Step 1: Upload Documents

• 

  Step 2: Processing Documents

• 

  Step 3: Training Data Pipeline (Dataset statistics)

• 

  Step 3: Training Data Pipeline (Edit dataset)

• 

  Step 4: Choose Output Style (Top)

• 

  Step 4: Choose Output Style (Bottom)

• 

  Step 5: AI Content Generation

• 

  Step 6: Model Training and Fine-Tuning (Top)

• 

  Step 6: Model Training and Fine-Tuning (Bottom)

🧠 AI Document Fine-Tuning Assistant

📌 About the Project

This project was born out of a desire to simplify and democratize how we fine-tune AI models on personal or organizational documents. The goal was simple yet ambitious: enable anyone to upload unstructured files—PDFs, DOCX, MD, TXT—and train an AI model that not only understands their content, but generates new material in their style and tone.

The unique part? The entire system was built from a single prompt using the AI agent bolt.new, showcasing what’s possible when code generation, planning, and deployment are automated.

🛠️ How It Was Built

• Data Parsing: Using LlamaParse and OCR where needed, documents are transformed into clean text for training.
• Dataset Generation: Automatically formatted into instruction–response pairs or document-style templates.
• Model Training: A quantized, fine-tuned LLM (e.g. Mistral-7B or LLaMA-2-7B) trained via QLoRA using free-tier GPU resources.
• Deployment: Frontend (Next.js + Vercel AI SDK) and backend (API endpoints) hosted on Netlify; GitHub used for version control.
• Execution: Fully bootstrapped from a prompt using bolt.new, making it replicable and accessible.

⚠️ Challenges

• Prompt Planning: The main challenge was condensing multiple requirements—fine-tuning, parsing, UI, deployment—into a single, comprehensive prompt for bolt.new.
• Tool Interfacing: Ensuring libraries and services worked smoothly together, especially across languages (Python for training, JS for frontend/backend).
• Maintaining Free-tier Constraints: Balancing performance with the limitations of free compute and hosting.

🌟 Outcome

This project demonstrates how anyone can deploy a powerful, fine-tuned AI that transforms unstructured content into smart, context-aware outputs like essays, presentations, and speeches—without writing code manually or paying a cent.

🌟 Disclaimer

Due to both time constraints and difficulties with deployment and a lack of credits due to the nature of the one-shot challenge, this participant did not have the time to put the Bolt Badge in time, this participant hopes the judge sincerely considers this case personally and will still allow this submission to be judged.

Built With

• express.js
• github
• javascript
• llama
• mistral
• netlify
• node.js
• python
• react