Show HN: Time-to-text time-series AI to limitate LLM hallucinations (github.com)

A new open-source framework has been introduced that combines deep learning methods and mathematical feature attribution to tackle hallucinations seen in traditional Large Language Models (LLMs) when generating automated financial reports. This hybrid local framework utilizes a unique architecture consisting of Temporal Convolutional Networks (TCN), Long Short-Term Memory (LSTM) networks, and Attention mechanisms, coupled with Captum Integrated Gradients to accurately analyze multi-asset time-series data. By redefining the way input data is processed—feeding the model with deterministic attributes—the framework ensures that language outputs are based on concrete computations, enhancing the reliability of the insights. This development is significant for the AI/ML community as it addresses a persistent issue where LLMs struggle with raw mathematical accuracy, particularly in contexts requiring nuanced time-series analysis. The architecture operates 100% on-device, maintaining data privacy, as the raw data remains on the user's machine. Only the final processed numerical weights are sent to open-source LLMs like Hugging Face for generating straightforward human-readable reports. This approach ensures an unbiased, mathematical perspective on financial data while empowering users to maintain control over their data and computational processes.