Forward-deployed / applied AI engineer and solutions architect, Minneapolis, MN. I build the knowledge layer for AI agents: Compressed Knowledge Graphs (CKGs) with typed, declared edges and a source URL plus SHA-256 hash on every node. The model handles language; the graph handles facts. The graph doesn't guess, it traverses. Patent pending.

LinkedIn · Book a call · daniel.yarmoluk@gmail.com · graphifymd.com · Hugging Face

Everything below resolves today.

• langchain-ckg - a LangChain retriever (pip install langchain-ckg), listed in the official LangChain integrations docs. 11 agent-stack graphs ship in the wheel and run offline.
• ckg-nvidia-ai - the NVIDIA AI developer stack as 20 graphs (1,006 nodes), served over MCP. Live demo: Hugging Face Space.
• glp1-pa-copilot (source) - healthcare prior-authorization draft gate on synthetic data. The model cannot add a fact: it verbalizes what the policy graph returns, a named reviewer accepts or edits, and every action lands in an append-only audit log.
• ckg-strands-showcase - CKG with Strands Agents: a one-model-call cap and a 55-test suite. The repo is an overview with reported results; the implementation is private.
• Hosted MCP endpoint at https://www.graphifymd.com/api/mcp (list_domains, query_ckg, get_prerequisites, search_concepts, query_intersect), plus a free library of graphs you can download as Markdown at graphifymd.com. The served library is 317 domain graphs (284 free, 33 Pro), 57,705 nodes, counted from the served files on 2026-10-06.
• Six production MCP services (design, auth, license gating, metering, telemetry, operations) and 80+ Claude Code skills, some public in skills-1.

Try an agent card that states its own economics, including when not to call it:

curl -s https://ckg-nvidia-ai.onrender.com/.well-known/agent-card.json | jq .economics

I treat evaluation as the product. Frozen corpora, SHA-256 manifests, paired runs, controls. I publish null results and limits: a LongMemEval-V2 run reported as the tie it was, graph-relative scores labeled as graph-relative, and a correction to my own earlier cost figures.

The one benchmark I quote, from one locked run (v0.6.2), 44 hand-curated domains, 7,758 queries:

That is nearly 4x RAG on structural queries. The 11x token reduction comes from the same run, not an independent result. At 5 hops, CKG F1 is 0.772 against 0.170 for RAG. Dataset: huggingface.co/datasets/danyarm/ckg-benchmark.

• Slalom - Solution Owner / AI and Digital Transformation Architect: UnitedHealthcare, Best Buy, Cargill. Discovery, build-versus-buy, production-readiness and evaluation requirements.
• West Monroe - Lead Solution Architect.
• ATEK Access Technologies - built the IoT and data science practice from zero.
• University of St. Thomas - Adjunct Professor, Graduate AI Systems (2018-present).

Fifteen-plus years of enterprise delivery.

Forward-deployed, applied AI, and solutions architecture roles where agents must be grounded, auditable, and deployed into real enterprise environments. Remote US; open to relocation.

LinkedIn · cal.com/daniel-yarmoluk-sjmnub · daniel.yarmoluk@gmail.com