Research

A DGX Spark evaluation of 13 locally run language models across a 21-task administrative work suite, covering calendaring, document parsing, financial and time-tracking calculations, email triage, and professional writing.

Which model should you put in your agent pipeline? We evaluate 11 frontier models across depth of reasoning, long-horizon reliability, tool-use accuracy, and engineering throughput, mapping each to concrete production use cases with price as a hard constraint.

LLMs quietly drift from their instructions as conversations grow longer. We show that dynamic control architectures significantly outperform standard prompting at keeping them on track.

Controlling emotional prosody in TTS without predefined labels is an open challenge. We combine StyleTTS 2 embeddings with HDBSCAN clustering on multi-modal acoustic features, achieving 200% more emotion clusters and 21.4% less unclustered data versus the baseline.

When laws change, RAG pipelines hallucinate outdated answers. We ablate retrieval components to find which combinations keep legal AI systems factually grounded over time.

Generative search is replacing ranked link lists with synthesised responses. We analyse how ChatGPT, Perplexity, Copilot, and Claude retrieve and surface information, and derive optimisation strategies with custom tools for measuring and improving AI visibility.

Identifying kidney stone composition guides treatment — but rare stone types strain neural-network methods. We show that combining NMF dimensionality reduction with a Gaussian Mixture Model on FTIR spectra achieves 75% concordance even with limited samples.

Real-time collision detection demands low latency at scale. We propose a field-edge architecture with batch processing and parallel threads that keeps YOLOv8-based vehicle detection above 30 FPS without sacrificing accuracy.

Low-resource languages with high grapheme-phoneme correspondence pose unique challenges for neural TTS. We demonstrate that training Tacotron 2 on a stressed Lithuanian corpus significantly improves speech naturalness as measured by MOS.