AI Solutions Architect · Forward Deployed Engineer · Minneapolis, MN
I work on context architecture — the layer between an agent and everything it is expected to know. Context routing decides which knowledge an agent gets and what it costs to get it. The knowledge layer underneath is typed and traversable, so every answer traces to a source instead of being reconstructed from similarity.
Fifteen years of enterprise architecture, most recently at Slalom delivering Fortune 500 AI. I like the deployed half of the job: the customer's environment, their constraints, the failure modes that never show up in a demo.
Every MCP server I run publishes an A2A agent card that states its own economics, so a calling agent can decide whether invoking is worth it before it spends anything:
curl -s https://ckg-nvidia-ai.onrender.com/.well-known/agent-card.json | jq .economicsThat last field is the point: it tells you when not to call it. A card that claims savings at every price is one a good agent should distrust.
RAG chunks prose and retrieves by embedding similarity, which discards the relationships. A CKG stores relationships as typed, authored edges and traverses them. Every answer traces to a source URL and a SHA-256 of the bytes it was authored from.
Benchmarked against RAG and Microsoft GraphRAG — 44 domains, 7,758 queries, locked at v0.6.2:
| CKG | RAG | GraphRAG | |
|---|---|---|---|
| macro-F1 | 0.471 | 0.123 | 0.120 |
| tokens per query | 269 | 2,982 | — |
| F1 at 5-hop depth | 0.772 | 0.170 | — |
The last row matters most. The advantage grows with question complexity, because multi-hop composition is exactly where embedding methods are weakest.
→ Clone the benchmark and re-run it · Dataset on Hugging Face (CC-BY-4.0)
The repo includes a reconciliation document correcting my own published cost figures — an earlier version priced CKG and the baselines against different models, which inflated the ratio. Numbers I can't defend are worse than no numbers.
|
Model Context Protocol Tool and output schema design · JSON-RPC initialize handshake · streamable HTTP and SSE transport · session management · DNS-rebinding transport security · per-method metering · HTTP 402 payment rails · MCP-native observability Built, shipped and debugged in production. |
Agent-to-Agent Agent cards advertising skills, auth, payment terms and machine-readable economics · x402 / HTTP 402 · EIP-3009 signed authorizations · Coinbase CDP facilitator · Base settlement · ERC-8004 agent identity |
Framework-agnostic by protocol. The same servers register unchanged in Semantic Kernel, LangChain, LangGraph, AutoGen, CrewAI, Claude and Cursor — integration happens at the protocol layer, so framework choice stays the caller's decision.
# Microsoft Semantic Kernel consumes an MCP server directly — no bridging code
from semantic_kernel import Kernel
from semantic_kernel.connectors.mcp import MCPStreamableHttpPlugin
async with MCPStreamableHttpPlugin(
name="ckg", url="https://ckg-nvidia-ai.onrender.com/mcp"
) as plugin:
kernel = Kernel()
kernel.add_plugin(plugin, plugin_name="ckg") # 9 tools → kernel functions12 packages on PyPI · 100+ releases · 6 running as live MCP services
| package | serves |
|---|---|
| ckg-nvidia-ai | NVIDIA developer stack — 20 domains, metered free tier, x402 payment challenge |
| ckg-nvidia-nemoclaw | NemoClaw stack — typed traversal with per-node provenance |
| ckg-agentforce | Salesforce Agentforce — license-gated tool surface |
| langchain-ckg | LangChain retriever — API-key auth, 402 handling |
| ckg-ai-platforms · ckg-nemotron-perplexity · ckg-agent-protocols | domain graphs |
Also here
- zep-ckg — Graphiti (Zep) plus CKG as a two-layer context agent
- Agent Skills — public Claude Code skills
Fifteen years of enterprise architecture. Fortune 500 AI delivery at Slalom across healthcare, retail and supply chain, including production-readiness and evaluation frameworks for HIPAA-regulated environments. Earlier: industrial IoT and commercial AI architecture at West Monroe, and a data science and IoT practice built from zero at ATEK.
Adjunct professor, University of St. Thomas — Graduate AI Systems. Featured in CIO Dive. Patent pending.
Open to AI Solutions Architect, Forward Deployed Engineer and Agentic AI Architect roles.
{ "price_usd_per_call": 0.010, "mean_tokens_returned": 269, "baseline_mean_tokens": 2982, // RAG over the same corpus "tokens_saved_per_call": 2713, "breakeven_input_price_usd_per_mtok": 3.69, "answer_quality_macro_f1": 0.471, "baseline_macro_f1": 0.123, "decision_rule": "…pays for itself on token cost alone when your input price exceeds $3.69 per million tokens. Below that, invoke only when answer quality matters." }