Turn ambitious ideas into working AI.

Organizations struggle with information overload and knowledge management. Valuable insights are buried in documents, emails, and other unstructured data sources, making it difficult to access and utilize institutional knowledge effectively.

We developed an AI Knowledge Automation system that extracts, organizes, and makes accessible the knowledge hidden in various data sources. The system uses natural language processing to understand context and semantic relationships between information pieces.

R&D projects generate vast amounts of information across multiple sources and formats. Project managers struggle to keep track of research findings, maintain context across team changes, and efficiently retrieve relevant information when making decisions.

We developed a Retrieval-Augmented Generation system specifically for R&D project management. It integrates with existing project tools to index, retrieve, and synthesize information from research notes, experimental results, papers, and communication channels.

Traditional machine learning requires centralizing data, which raises privacy concerns and can be impractical for sensitive domains such as healthcare and finance. Organizations need a way to train models collaboratively without sharing raw data.

We implemented a federated learning system using the Flower library that allows multiple parties to train a shared model while keeping their data private. The platform supports distributed datasets, custom aggregation, unreliable client availability, and enhanced privacy controls.

Pokémon cards have become valuable collectibles, with rare cards often being counterfeited. Authenticating cards and identifying their value requires expertise and can be time-consuming.

We developed a computer-vision system that automatically detects and classifies Pokémon cards from images and video. It uses deep-learning techniques to recognize card features, identify the character, determine rarity, detect visual counterfeit patterns, and connect results with pricing data.

Haitian Creole is an under-resourced language with limited digital accessibility tools. This creates barriers for Haitian speakers accessing technology and services that depend on reliable speech recognition.

We developed a speech-to-text system specifically for Haitian Creole using transfer learning and limited training data. The system adapts pre-trained models to Haitian Creole phonetic structures and vocabulary while supporting dialect variation, noisy environments, and mobile-friendly deployment.