
AUDITIA is a high-performance AI auditing agent engineered to combat opacity in public contracting. Traditional rule-based systems often fail to catch complex, multifaceted corruption. To solve this, we built an unsupervised Autoencoder using Google's JAX and Flax frameworks. The system ingests real Colombian contracts from the SECOP II API. It normalizes financial data (value and time) via Z-scores, applies One-Hot Encoding to categorical data (like departments), and processes free-text descriptions using a local HuggingFace MiniLM Transformer. This creates a precise 460-dimensional mathematical vector for each contract. By leveraging the AMD Instinct MI300X and the ROCm software platform , the JAX XLA compiler translates our model directly into optimized MIOpen instructions. The Autoencoder is trained on historical data to reconstruct "normal" contracts. When an irregular contract is processed, the model fails to reconstruct it; this Mean Squared Error (MSE) serves as an objective Anomaly Score. A high score automatically triggers an integrated Generative LLM agent (Gemma) to explain the specific red flags found. This architecture allows us to process millions of public records at extreme speeds entirely locally, eliminating external API costs and maximizing the massive memory bandwidth of AMD GPUs.
10 May 2026