Category
Explore by categories
Identifies and surgically removes the internal activation directions that trigger refusal behavior in large language models, with one-click options on a HuggingFace Space or a local Python API. Combines multiple extraction methods (SVD, whitened SVD, sparse autoencoders), reversible steering, and analysis-informed verification to quantify capability and refusal trade-offs.
Visualizes live global data on a CesiumJS 3D globe via a plugin-driven pipeline and real-time WebSocket DataBus. Supports dynamic plugin marketplace, an opt-in Agent Bus for external LLM/MCP control, and self-hosting with Docker and PostgreSQL.
Lets an LLM autonomously propose, edit, run, and evaluate short single‑GPU LLM training experiments — fixed 5‑minute runs (~12 experiments/hour). Agent edits a single train.py; humans supply goals via program.md. Single‑GPU, val_bpb metric.
Measures multiline text layout and block height without triggering browser reflow: it measures text segments once via Canvas+Intl.Segmenter and caches widths, then computes line breaks with pure arithmetic. Useful for streaming AI text, virtualization, and custom per-line rendering.
Provides a single persistent database and open protocol so multiple AI tools share the same memory — built-in vector search, an AI gateway, and capture/skill extensions. Best for teams and power users who want a unified, self-hosted agent memory instead of siloed notes or per-tool caches.
Turns a single research idea into runnable experiments and a conference-ready paper by orchestrating an LLM-driven end-to-end workflow (literature → design → code → sandboxed runs → analysis → writing). Provides human-in-the-loop checkpoints, domain-specialist executors, and multi-layer citation verification.
Provides a persistent, typed semantic memory layer for AI agents—supports remember, recall, and answer primitives so agents retain long-term context. Writes are instantly searchable and retrieval uses an information-theoretic engine, avoiding separate vector DBs or indexing delays.
Compresses high-dimensional embeddings into low-bit TurboQuant indexes for fast, memory-efficient local vector search. Supports online ingest (no train/rebuild), SIMD kernels that match or beat FAISS, per-vector length-renormalization, and runtime allowlists — suited for privacy-sensitive, low-latency RAG.
Turns a repo's code, docs, PDFs, images, and videos into a queryable multimodal knowledge graph for AI coding assistants. Uses deterministic AST extraction for code and LLM-based semantic extraction for other assets, exporting interactive HTML, JSON, and a human-readable audit report.
Provides a cloud-backed shared memory and skill-propagation layer for coding agents: captures session traces, mines recurring patterns into reusable SKILL.md, and shares capabilities across agents in real time. Features hybrid semantic+lexical search, BYOC storage, and a VFS for traces — built for team workflows and agent orchestration.
Compiles raw documents into a persistent, interlinked Markdown wiki that LLMs can query; uses PageIndex for vectorless, reasoning-based retrieval of long documents, supports native multi-modality, bundled web Workbench, and skill distillation.
Provides a brain layer for AI agents that synthesizes answers, traverses a self-wiring knowledge graph, and highlights gaps in team knowledge. Ships hybrid retrieval, citation-aware synthesis, and MCP integrations for Claude/Codex to power meeting prep and company-wide memory.