23+ years building software — from COBOL and Smalltalk to Rust, Agentic Memory, and Sovereign AI

I work on agentic memory architectures, world models for AI agents, and secure/sovereign agents — all grounded in embedded graph databases and edge-first, privacy-first principles.

• Agentic Memory — Multi-layered memory architectures (episodic, semantic, procedural) built on knowledge graphs. Cognitive extraction-consolidation pipelines. Memory that supports causal reasoning, not just retrieval
• World Models for AI Agents — Structured representations of the environment an agent reasons about to make decisions
• Secure & Sovereign Agents — Self-sovereign identity (DIDs, KERI, Verifiable Credentials), agent-owned credentials, trust without centralized authority
• Edge AI & Embedded Graphs — Local-first, privacy-first. Graph databases that ship as a single file and run in-process. No server, no network hop, no data leaving the device

• oxilite — Oxigraph-compatible RDF database and SPARQL 1.1 engine that runs on plain SQLite instead of RocksDB, including Cloudflare D1 — same data model, same SPARQL semantics, same Rust API, with a compiled SPARQL→SQL planner, RDFS/OWL reasoning, SHACL/ShEx validation, and openCypher over the same data. Dual MIT/Apache-2.0 — oxilitedb.com · crates.io · source on GitHub
• oxilite studio — VS Code development studio for oxilite: live SPARQL, Datalog and Cypher with completion and go-to-definition, RDFS/OWL reasoning with proof explanations, SHACL/ShEx validation in the Problems panel, a store explorer, notebooks, and an MCP server for agents. Dual MIT/Apache-2.0 — source on GitHub
• Factum — Object-Role Modeling (ORM 2) as a VS Code extension. Draw a conceptual schema, read it back as plain-language sentences a domain expert can confirm, and map it to relational (SQL) or property graph (LadybugDB) schemas. Ships an MCP server (factum-mcp) and a CLI (validate, verbalize, diff, drift, derive) so coding agents read the conceptual model instead of guessing it from column names. MIT licensed — source on GitHub
• LPG Modeler — VS Code extension for labeled property graph schemas. Author one YAML model on a canvas (with proper inheritance vs. mixins) and generate LadybugDB DDL, Neo4j constraints, SHACL shapes, and an OWL ontology from it — with everything a target can't enforce reported as a diagnostic instead of silently dropped. Checkable in CI. MIT licensed — source on GitHub
• Causal Canvas — Visual editor for causal models (DAGs, ADMGs, PAGs, causal loop diagrams) backed by a JSON-native format (CausalJSON) and a CLI. Includes a causal linter that catches collider adjustment, invalid instruments, and unidentifiable latents, plus reproducible SVG/PDF figure generation for CI. Fully local, no telemetry. Apache-2.0 — source on GitHub
• Hybrid Graph RAG (Rust) — Four retrieval modes in one query: vector search + graph traversal + PageRank + community detection. +109% on multi-hop questions vs vector-only RAG
• Hybrid Graph RAG (Python) — Reference implementation of Hybrid GraphRAG with LadybugDB

I write the Age AI book series on Leanpub:

🇺🇦 I also write a Ukrainian children's book series: Макс, Цугі і світло, яке чутно

Primary: Rust, TypeScript, React, JavaScript
JVM: Kotlin, Scala, Java, Clojure
Systems & Crypto: Rust, Zig, C
Formative: Smalltalk (Pharo, Cuis, Self), COBOL — these shaped how I think about objects, agents, and message-passing

• Substack — Sovereign Agentic AI — deep dives on agentic memory, world models, LadybugDB, Hybrid RAG
• Medium — technical articles
• YouTube — video content

🌐 https://www.pavlyshyn.me

Nostr: npub1u6qhg5ucu3xza4nlz94q90y720tr6l09avnq8y3yfp5qrv9v8sus3tnd7t

The next generation of AI will not simply compute. It will remember.