Each project now lives on its own subdomain of emmaleonhart.com — consolidating what used to be scattered across separate domains into one place.
A geometrically compiled language where logical operations over vector spaces are resolved at compile time into matrix multiplications. Branches are continuous weighted blends, loops are geometric rotations, and programs compile to straight-line tensor work — GPU-native and differentiable by construction.
A lean, high-performance RDF-star triplestore written in Rust with native HNSW vector indexing, ontochronological temporal queries, and a SPARQL+ query language — a single system where vectors are just triples, replacing both a vector database and a SPARQL triplestore.
A pretrained transformer given a genuine time axis: a fixed, randomly-initialized reservoir is injected into a mid-layer's attention so the model carries state across separate forward passes. Attention runs jointly over token hidden states and reservoir nodes, the reservoir reads from the attention matrix through frozen weights and writes its state back through a learned readout each pass, and only the upper layers are fine-tuned, so recurrence comes from untrained reservoir dynamics rather than learned memory.
A beginner-friendly guide to vibe coding as a discipline — building real software through AI collaboration while staying in control, with documentation, version control, and tests. Nine chapters from a zero-experience setup to a working AI-assisted workflow, scaffolded with cleanvibe.
Reproducibility artifact for the Claw4S 2026 paper applying TransE-style relational displacement analysis to frozen embedding models with Wikidata triples as probes — surfacing 30 model-agnostic relational operations and a silent [UNK] tokenizer defect in mxbai-embed-large.
A zero-dependency Python CLI that scaffolds AI-assisted coding projects and launches Claude Code — injecting an opinionated CLAUDE.md contract plus queue.md/todo.md/devlog.md so a project starts structured, not as a blank-repo scramble (done = delete from queue.md + a dated devlog entry, same commit). Stable v1.2.0: new (now prompts on a non-empty dir, inits on main), clone, convert, replicate.
A social network you run locally from your own desktop. QueryKey has the elements of a personal relationship manager and uses local AI agents to help you keep up with the people in your life — while respecting your privacy and theirs. Built on a local-first ingest → knowledge-graph engine.
Research testing whether redemption-narrative system prompts measurably move emergently misaligned LLMs back toward alignment — behaviorally, on self-rated harmfulness, and geometrically against a derived misalignment direction. Theoretical framing: emergent misalignment as moral injury, with cross-scale results across Llama and Qwen.