CLOUDWALK AI

We explore transformers and other deep architectures beyond traditional language models, pushing the boundaries of applications in financial technology. Our experiments include:

‍

1. LLMs in credit assessment: Analyzing unstructured data sources to enhance credit decisioning.
2. Fraud detection: Leveraging language models to identify suspicious patterns in transaction descriptions and customer communications.
3. Merchant Vector Space: A novel approach where we build embeddings to describe merchants rather than conventional word tokens, enabling non standard analysis of business relationships, behaviors and information retrieval.

We aim to contribute to the broader field of AI research and prepare for the potential emergence of Artificial General Intelligence (AGI). Our Consciousness Project is a forward-looking initiative by means of which we navigate the limits of AI capabilities. That includes giving eyes, ears, and free will to a Large Language Model. Key aspects of this project include:

‍

1. Sensory integration: Developing methods to incorporate diverse sensorial inputs into language model processing
2. Agency simulation: Exploring decision-making capabilities of Large Language Models (LLMs) through game-playing scenarios and Investigating the potential for autonomous behavior and strategic thinking in AI systems
3. Subjective experience modeling: Investigating ways to replicate or simulate human-like mental processes and consciousness.

We leverage cutting-edge Large Language Model (LLM) agents to enhance customer experience. Our AI assistants handle a substantial portion of chats without human intervention, combining LLMs with advanced techniques:

‍

1. Workflow Graphs: We use structured representations of typical interaction patterns to guide our AI agents through conversations, enhancing their ability to navigate complex customer inquiries effectively.
2. Retrieval Augmented Generation: This allows our agents to access and incorporate relevant information from our knowledge base in real-time, providing accurate and contextually appropriate responses.
3. Dual-Agent System: We use a main conversational agent and a specialized tool agent working together. The tool agent can perform specific actions typically done by human analysts, allowing for faster and more efficient support.

We continually refine these systems, focusing on improving empathy, problem-solving capabilities, and seamless escalation to human agents when necessary.

We harness the power of evolutionary algorithms for feature selection and hyperparameter optimization. Inspired by natural selection we treat features and parameters as genes in a genetic pool, we evolve model "species" over hundreds of generations. This approach allows us to efficiently explore vast solution spaces and discover optimal feature sets and hyperparameters for each specific problem. 
‍

1. Initial population: We start with a diverse set of model configurations.
2. Fitness evaluation: Each model is tested on our dataset, with performance metrics serving as fitness scores.
3. Selection: High-performing models are selected to "breed" and produce offspring.
4. Crossover and mutation: New models are created by combining features of successful parents and introducing random and strategically targeted mutations.
5. Iteration: This process repeats over many generations, gradually improving the overall population.

Our proprietary models go beyond standard underwriting methods, leveraging actual customers behavior to assess their repayment capabilities.

As effective scoring must be dynamic, we often evaluate all of our customer base and re-train our models, ensuring that the assigned score accurately reflects the most up-to-date data we have. Our behavior analysis takes into consideration patterns across all contact points between us and our customers, leading to a comprehensive understanding of creditworthiness, which proved to be more reliable than traditional decision engines, data sources or paradigms.

We also incorporate non-traditional data sources to assess credit risk, allowing us to evaluate and actually assign loans to people with little data available (such as "thin-file" and underbanked individuals).

We make use of Data Science techniques to map and classify user behavior throughout the client journey, enabling personalized experiences and targeted interventions:

‍

1. Customer Attraction: Using Marketing Mix Models (MMM) to optimize campaigns and generate quality leads. These models analyze the impact of various marketing channels, taking into consideration factors such as seasonality, competition, and external events, to allocate resources efficiently .
2. Lifetime Value (LTV): We categorize users and predict their long-term value using machine learning techniques. This involves analyzing historical data, transaction patterns, product usage, and engagement metrics to forecast future revenue and prioritize customer relationships.
3. Churn models: We develop churn detection models based on each client's unique transactional patterns. These models consider industry-specific factors, seasonal trends, and individual user behavior to predict and prevent customer churn.

Our fraud detection system is designed to minimize fraud rates while maintaining high transaction approval. We've implemented a comprehensive three-layer system that detects and prevents fraud in different perspectives:

‍

1. Transactional perspective: These ultra-fast models examine each transaction in real-time, deciding within milliseconds whether or not it should be approved. They analyze numerous features such as transaction amount, merchant category, time of the day and geographical location, to make split-second decisions.
2. Behavior perspective: By analyzing sequences of events over time, these models distinguish between fraudulent and normal user behavior. They have more flexibility as to the response time, allowing for deeper analysis of patterns such as spending habits, device usage and account activity trends.
3. Relational perspective: These models focus on detecting group-based fraud by mapping relationships between users and identifying suspicious graph topologies. By examining connections between accounts, devices and transactions, we can uncover complex fraud rings and coordinated attacks which might remain unnoticed when applying  traditional detection methods.

Our Geospatial system takes advantage of various location-based data, giving us valuable insights about the global diverse landscape of where we operate and also allow us understand each client individual behavior by looking at a dispersion analysis.
‍
By integrating geospatial intelligence into our AI ecosystem, we enhance the relevance of our products, and can develop tailored financial solutions to our customers. This data-driven approach allows us to adapt to regional economic shifts, capitalize on emerging opportunities and provide hyper-localized solutions across Brazil's diverse geographical landscape.
‍
Geospatial data and map visualizations are made available for different teams in CloudWalk, allowing us, for example, to set demographic profiling for targeted marketing and tailor services or product to specific local market needs.

Our AI-driven monitoring system tracks hundreds of time-series metrics and KPIs, critical for keeping our business operations smooth. Key features include anomaly detection, forecasting and setting up adaptive thresholds. Our models learn from extensive historical data to accurately predict and adjust for seasonal patterns, holidays, and special events.

We prioritize high-quality data over complex model architectures, which forms the foundation of our AI initiatives. We focus on three core areas:

‍

1. Robust feature market: A centralized platform for creating, sharing and managing features across the organization, promoting reusability and consistency in model development.
2. Advanced MLOps systems: Streamlined processes for model development, deployment and monitoring, ensuring rapid iteration and maintaining model performance in production.
3. Cutting-edge data infrastructure: Scalable and efficient systems for data storage, processing, and analysis, supporting real-time and batch operations.

We continuously optimize data storage, quality, and accessibility, developing tools for anomaly detection and cost management. Our cross-functional approach spans from creating analytics solutions to supporting AI initiatives and managing critical database operations.