AI Research Engineering · London
Lorenzo Moriondo
I build graph-wiring methods, spectral vector search, and agentic code-generation workflows — and ship them as open-source Rust and Python.
From Embedding Geometry to Spectral Search: Energy Dispersion Networks for Vector Retrieval
The Graph Wiring framework: exploiting feature-space spectral structure in vector search, with Spectral Indexing as its task-specific instantiation. Couples geometric similarity with spectral information to improve head–tail coherence and semantic alignment beyond purely geometric retrieval, supports adaptive search via tau-modulation, and grounds its theory in epiplexity. Evaluated across benchmark and industrial settings with the open-source arrowspace library. With Ilias Azizi.
Graph Wiring: Accessing More Information via Feature-Space
A constructive separation and conditional coding analysis: two corpora with identical candidate–query row Gram matrices but different column graphs, and a convex blend of cosine and Rayleigh-energy proximity that exploits the difference. The blend reaches Top-1 accuracy 7/8 against 1/2 for every observer restricted to the complete row Gram, with an optimal conditional log-loss reduction of 1−h₂(1/8) bits.
Cosine Is Not Semantics: Learning a Spectral Metric from the Embedding Corpus
A corpus-conditioned alternative to cosine: a graph over embedding features supplies a positive-semidefinite operator that changes the geometry in which angular agreement is measured. The resulting normalized spectral kernel compares graph gradients of embedding signals, admits an explicit cut-invariant feature map, and recovers ordinary cosine under the identity operator.
ALD-SC: A Spectral Latent Diffusion Model
A spectral latent diffusion model where decoding runs on the feature-space manifold defined by a frozen ArrowSpace graph Laplacian. A graph-structured decoder replaces unconstrained convolutions with propagation along the Laplacian's smooth eigenvectors, gated by a Barontini entropic clock that terminates sampling intrinsically.
ArrowSpace: Spectral Search For Embeddings
Spectral indexing for vector similarity search combining cosine similarity with graph-Laplacian roughness (λτ-indexing) to produce topology-aware similarity scores. Published in the Journal of Open Source Software.
Vibrational Deduction Transformer
A transformer architecture that reasons through vibrational (oscillatory) modes in embedding spaces, using spectral decomposition to uncover latent structure beyond static attention patterns.
Graph Wiring: Vector Analysis & Retrieval
A library for vector analysis, retrieval and curation using graph-wiring techniques. Explores how topological structure in embedding spaces improves retrieval quality beyond naive cosine similarity.
Epiplexity: Spectral Feature Interactions in Embeddings
Epiplexity measures spectral feature interactions in high-dimensional embedding spaces, providing a formal framework for understanding how graph Laplacian eigenstructure influences retrieval quality and manifold geometry.
