Ironhack AI

During IronHack AI Engineering Bootcamp, I had to work on several projects.

For all of them I had a week to code and solve the subject.

Challenge

For this project, I had to create a chatbot. I built an AI-powered React component generation service that transforms text descriptions and UI mockup images into production-ready React code using OpenAI's GPT-4 and Vision models.

Key Features Implemented

• Smart Code Generation: Natural language to React/TypeScript components with Tailwind CSS styling by default.
• Image Analysis: Upload UI mockups and get corresponding React code using OpenAI Vision API.
• Contextual Learning: Vector similarity search via Pinecone database for relevant code examples.
• Interactive Chat Interface: Real-time conversation-based development workflow with session management.
• Automated Knowledge Base: Population from HuggingFace datasets and custom code snippets.

Technical Architecture

• Backend: Flask API with OpenAI integration, MongoDB for storage, Pinecone for vector search, Cloudinary for image processing.
• Frontend: React/TypeScript with Vite, featuring drag-drop image upload, real-time messaging, and code syntax highlighting.
• AI Pipeline: LangChain orchestration with LangSmith monitoring, custom prompt engineering for React best practices.

Special Features

Automatic detection of boilerplate requests redirects users to my custom CLI tool (@julseb-lib/julseb-cli), showcasing a complete AI-powered development ecosystem.

See the full project on Github.

Project to classify images to certain categories, using CIFAR-10 dataset to train the model.

Challenge

For this project, I had to work on data preprocessing, CNN architecture, optimisation techniques, transfer learning, and model evaluation.

Preprocessing steps

After loading the data and plotting a matrix of 10 x 10 images, the data was preprocessed with:

• One-hot-encoding in order to transform categorical data into numerical format
• Normalization in order to scale the data between 0 - 1 so that any feature does not disproportionately influence the algorithm
• Data Augmentation in order to increase the diversity and size of a training dataset by creating modified copies of existing data. This is to help machine learning models generalize better and reduces overfitting.

Training Process Details

• All models compile using Adam optimizer, categorical loss entropy, and accuracy metrics. Each model modifies the learning rate to see how it effects the accuracy of the model.
• I used ReduceLROnPlateau to reduce the learning rate when validation loss plataeus
• I used early stopping to prevent overfitting by halting training when the model's performance on a validation set starts to decline.

See the full project on Github.

A Natural Language Processing project that reads, understands, and summarizes product reviews — so you don’t have to. This AI-powered tool classifies review sentiments, groups similar opinions, and generates short summaries to save users time and help businesses make smarter decisions.

Features

• Data Preprocessing: clean and prepare review text by removing noise, filtering, and tokenizing.
• Sentiment Classification: uses a fine-tuned RoBERTa model to classify reviews as Negative, Neutral, or Positive.
• Clustering: groups similar reviews using KMeans, then visualizes the clusters using PCA and t-SNE.
• Summarization: summarizes review clusters using BART and T5 models for quick insights.
• Interactive App: a simple Streamlit interface where you can input a review or product name and see:
• Sentiment prediction
• Summary of grouped reviews
• Cluster visualization

Models Used

• cardiffnlp/twitter-roberta-base-sentiment – For sentiment classification
• facebook/bart-large-cnn – For review summarization
• t5-small – Alternative summarization model for concise outputs

Results

• Classify reviews into Positive, Neutral, Negative
• Group similar reviews using clustering
• Generate short summaries for each group
• Visualize insights with PCA and t-SNE
• Simple, user-friendly interface via Streamlit

See the full project on Github.