This is the Pexon engineering blog: every post belongs to one business area hub and is also listed here by date. The hubs collect posts on one subject: running open-weight models on hardware you control, rolling Claude out across an enterprise, and making SAP, MES and PLM data reachable for AI systems.
68 posts
Claude and Anthropic · cost
Claude API credits are a monthly Console balance on Max and Team: $100 or $200 on Max, $20 or $100 per Team seat to a $500 cap. Unused credit expires.
10 min read · updated 29 Oct 2026
Claude and Anthropic · mcp
Hosted MCP server or self-hosted? Makini, CData Connect AI, vendor servers and custom builds compared on identity, KRITIS, write governance and upkeep.
9 min read · Phillip Pham
Private AI · inference performance
KV cache size grows linearly with context: 5 GiB at 32k tokens for Mistral NeMo 12B, measured. The n² attention matrix is not what fills a 12 GB card.
8 min read · Flavius Moldovan
Private AI · inference performance
LLM batching measured on one 12B model: 28 tokens/s for one user, 194 for 32 together, each of them 4.7 times slower. Where the trade turns.
8 min read · Flavius Moldovan
Private AI · inference performance
Prefill vs decode, measured on one 12B model: 284 tokens/s reading the prompt, 29 tokens/s writing the answer. Decode is bound by memory bandwidth.
9 min read · Flavius Moldovan
Claude and Anthropic · mcp
MES integration with AI agents for batch release and deviation triage across MES, LIMS and QMS: the architecture, where it breaks, what GxP leaves to people.
10 min read · Phillip Pham
Claude and Anthropic · mcp
The Windchill REST API can serve AI agents, but four things break a naive MCP server: one shared account, default config specs, POST reads and mixed write tools.
11 min read · Phillip Pham
Data foundation · source systems
An energy market message on RAIDA moves by email, SFTP or REST. The assistant may read the status. It does not sign, encrypt or send the XML.
5 min read · Phillip Pham
Data foundation · source systems
A grid control archive is the history PSIcontrol, PSIprins, Spectrum Power 5 or HIGH-LEIT already keeps. The assistant may read it. A setpoint is not a question.
5 min read · Phillip Pham
Data foundation · source systems
A meter data gateway holds the values BTC, co.met or GWAdriga already forward. The assistant may read a value. Gateway administration is a different right.
5 min read · Phillip Pham
Claude and Anthropic · mcp
The Teamcenter API for AI agents: which surface to build on, the five calls behind one BOM read, and four traps to close before an MCP server goes near production.
9 min read · Phillip Pham
Data foundation · source systems
CRM quote questions an agent may answer: which quotation is open, its status, the customer. Accepting it, or turning it into an order, stays a click in the CRM.
5 min read · Phillip Pham
Data foundation · source systems
CRM sales questions with a documented read: open deals, the owner, the next date. Stage changes, sends and deletes are not questions. They are edits.
5 min read · Phillip Pham
Data foundation · source systems
The DATEV API splits in two: DATEVconnect on the workstation for account postings, and the cloud gateway for accounting sequences. An agent may read, not post.
6 min read · Phillip Pham
Data foundation · source systems
Utility billing integration for a Stadtwerk: the agent reads a contract on Schleupen.CS under the caller's login. Market messages stay in the billing system.
9 min read · Phillip Pham
Data foundation · source systems
Energy data management at a Stadtwerk sits in BelVis, Robotron or SOPTIM. An agent may read a schedule status. Sending that schedule stays inside the system.
8 min read · Phillip Pham
Data foundation · source systems
Utility network documentation at a Stadtwerk lives in Smallworld, Lovion and SAP PM. An agent may ask which meters an outage covers. It does not edit the model.
8 min read · Phillip Pham
Private AI · assistants
A ChatGPT alternative by constraint: Claude for long work, Perplexity for citations, Mistral for EU hosting, or self-hosting when prompts must stay in-network.
13 min read · Phillip Pham
Claude and Anthropic · mcp
ERP PLM integration with an AI agent: two MCP servers, one eBOM-to-mBOM comparison tool, and four configuration mismatches that make an agent compare the wrong BOMs.
10 min read · Phillip Pham
Claude and Anthropic · mcp
Dynamics 365 MCP cost for Claude: 0.1 Copilot Credits per tool call to Microsoft, then the token bill, which is usually the larger line. Worked out in euros.
8 min read · Phillip Pham
Company brain · agentic ai
Custom engine agents bring your own orchestrator and models into Microsoft 365 Copilot. Stay declarative until the workflow leaves Copilot's loop.
6 min read · Phillip Pham
Company brain · agentic ai
Copilot Studio implementation needs Copilot Credits, a 125% stop rule, and four layers before any production channel opens — not a demo agent.
7 min read · Phillip Pham
Claude and Anthropic · mcp
Industrial MCP server security, checked against five vendor OT servers: which ship write tools, what the defaults are, and what a read-only deployment needs.
8 min read · Phillip Pham
Data foundation · mcp integration
MCP server for enterprise source systems: 100 niche ERP, MES and PLM systems in European industry with no AI connector. How Pexon builds the missing MCP servers.
6 min read · Noel Dinger
Private AI · workflow automation
n8n vs LangGraph for AI workflows: visual automation builder against agentic graph framework. When drag-and-drop wins and when code-first pays.
4 min read · Noel Dinger
Private AI · engineering tooling
AI tools for engineers that earn their place in 2026: coding agents, model gateways, observability and evaluation. A practical stack, honestly assessed.
5 min read · Noel Dinger
Private AI · open source interface
Open WebUI self-hosted: a private ChatGPT-style interface for your own models, with RAG, multi-user access and no data leaving your estate.
5 min read · Noel Dinger
Claude and Anthropic · claude code
Claude Code audit trail: a Claude API audit uses the Compliance API, not the Activity Feed. Transcripts stay six years. Bedrock, ZDR and ~/.claude stay yours.
9 min read · Phillip Pham · updated 6 Oct 2026
Claude and Anthropic · mcp
MCP servers connect AI assistants to CRM, ERP and databases through one protocol. The enterprise angle: architecture, registry, tool poisoning, and secure setup.
8 min read · Phillip Pham
Private AI · inference serving
Model Hardware Standard: Anthropic's 27 August 2026 research preview for agents on lab devices. MCP is one access path. Open source is the named successor.
5 min read · Noel Dinger · updated 14 Sept 2026
Data foundation · delivery model
Buying forward deployed engineering means testing whether the work lives on the vendor's platform or in your SAP, MES and PLM, not copying Palantir's GTM.
9 min read · Phillip Pham
Claude and Anthropic · cost
Send Claude the same long document twice: prompt caching writes it at 1.25x and reads it at 0.1x inside five minutes. Batch processing is billed at 50 percent.
15 min read · Noel Dinger · updated 29 Sept 2026
Company brain · rag
RAG, fine-tuning and prompt engineering solve different problems. The three-way matrix shows which fits current data, stable patterns or plain instructions.
7 min read · Phillip Pham
Claude and Anthropic · cost
Claude Code exports 8 OpenTelemetry metrics. What to instrument, how to attribute spend to a team, and which of those numbers will mislead your board.
15 min read · Phillip Pham
Company brain · rag
RAG system build guide: chunking strategies, embedding models, vector search, retrieval quality, evaluation, and the rollout that survives a security review.
9 min read · Phillip Pham
Company brain · document processing
Intelligent document processing: extract invoices, delivery notes and POs, validate against ERP master data, then human review on exceptions.
7 min read · Phillip Pham
Private AI · inference serving
llm-d is the Kubernetes-native stack above vLLM: cache-aware routing, optional prefill/decode split, founded by Red Hat, Google, IBM, CoreWeave and NVIDIA.
5 min read · Phillip Pham
Private AI · open weight serving
Nemotron is NVIDIA's open-weight family. Nemotron 3 Ultra is 550B/55B MoE, OpenMDW 1.1, self-hostable when prompts must stay on your GPUs.
5 min read · Phillip Pham
Claude and Anthropic · coding agents
Claude Code vs Codex on the same task: 11x faster, 6.6x cheaper for one, deeper testing and architecture for the other. Orchestration decides.
7 min read · Phillip Pham
Claude and Anthropic · coding agents
OpenCode is the open-source terminal coding agent with bring-your-own-model: custom agents, local models, and a real challenge to Claude Code.
9 min read · Phillip Pham
Claude and Anthropic · cost
GPT-5.6 Sol vs Claude Fable 5 in real work, not benchmarks: Fable is the manager, Sol the worker — and model routing is what decides your AI costs.
7 min read · Phillip Pham
Company brain · agentic ai
Prompt engineering is the recipe; agent engineering is the cook. The 7 skills — system design, tool contracts, retrieval, reliability, security, metrics.
6 min read · Phillip Pham
Company brain · agentic ai
AI workflows follow a fixed path a human defines; AI agents let the model decide. RAG is a workflow, ReAct is the agent pattern. When each is the right call.
6 min read · Phillip Pham
Company brain · agentic ai
Scaling agents is not about more infrastructure. Errors propagate, costs grow nonlinearly, and one agent owning everything becomes the bottleneck.
5 min read · Phillip Pham
Private AI · model routing
GLM 5.3 Flash: open-weight, 1M context, and an audit of 1,000+ PRs for 12 cents where a frontier model costs over $100. Routing just got cheaper.
6 min read · Phillip Pham
Private AI · agent orchestration
LangGraph structures agentic AI as a directed graph of nodes, edges and state — control, visibility and human-in-the-loop for production agents.
6 min read · Phillip Pham
Private AI · inference serving
LLM self hosting on Kubernetes is production reality: 650B tokens a month, 30 tenants, shared GPUs. The four optimizations that make it economical.
6 min read · Phillip Pham
Private AI · agent orchestration
The same model can shine in one harness and stall in another. Most recent gains came from the harness — tools, memory and the agentic loop.
6 min read · Phillip Pham
Private AI · open source platform
The open-source AI platform is the harness, not the model. How open harnesses, model routing and a 90/10 cost split keep your architecture independent.
7 min read · Phillip Pham
Private AI · llm gateway
What is LiteLLM? The open-source gateway that puts one API in front of every LLM provider — virtual keys, team budgets, usage tracking, no lock-in.
8 min read · Phillip Pham
Claude and Anthropic · eu hosting and gdpr
There is no GDPR switch in Claude. Fourteen checks decide whether a rollout is lawful — contract, processor, retention, transfers, records — and who owns each.
15 min read · Phillip Pham
Claude and Anthropic · coding agents
Vibe coding means describing software to an AI instead of writing it line by line. The opportunities, the limits, and the review that keeps it out of production.
7 min read · Phillip Pham
Claude and Anthropic · cost
Claude API cost: input, output and cache pricing per model, the token multiplier that catches teams, and how to forecast a bill from your actual traffic.
5 min read · Phillip Pham
Private AI · inference platform
Azure AI Foundry vs Kubernetes for LLM workloads: managed model serving against self-operated inference on AKS. Cost, control, GPU scaling and when each wins.
6 min read · Noel Dinger
Private AI · llm gateway
LiteLLM vs OpenRouter: self-hosted gateway with budgets and audit logging against hosted routing with one API key.
5 min read · Noel Dinger
Private AI · llm observability
MLflow vs Langfuse for LLM observability: experiment tracking and model registry against tracing, evals and production monitoring.
4 min read · Noel Dinger
Private AI · inference serving
Ollama vs vLLM for LLM serving: developer ergonomics against production throughput. When Ollama is enough, when vLLM pays, and how to run both behind one gateway.
4 min read · Noel Dinger
Private AI · open weight serving
Qwen hosting on your own hardware: model selection, vLLM vs Ollama setup, GPU sizing and EU data residency.
4 min read · Noel Dinger
Private AI · llm gateway
LiteLLM proxy: Docker first, virtual keys not the master key, budgets per team. Self-hosted for control, not cheaper tokens. Monitor before the first incident.
9 min read · Phillip Pham
Claude and Anthropic · claude code
Claude Code vs GitHub Copilot: Copilot wins central policy, Claude wins per-developer cost. Own tenant vs GitHub's, 10% residency premium, 180-day vs six-year audit.
21 min read · Phillip Pham
Claude and Anthropic · copilot
3-year TCO: copilot vs own ai platform — Copilot at ~€26/user/month vs. open AI platform at €0.02–0.05/user/day. Breakeven point, hidden costs, option comparison.
12 min read · Noel Dinger
Claude and Anthropic · claude code
Parallel terminals and night agents, not the licence, drive Claude Code pricing to 2,500 euros a month — about 1.5 senior engineers. How IT leaders budget it.
7 min read · Phillip Pham
Claude and Anthropic · cost
Claude token costs are not a stable unit: Anthropic documents that 4.7 and later produce about 30% more tokens for the same text. Compare per task instead.
5 min read · Noel Dinger
Claude and Anthropic · claude code
When not to use AI agents: four problem shapes where a script, a form or a person wins, and the one question that separates them from the cases agents suit.
5 min read · Phillip Pham
Claude and Anthropic · cost
Switching model vendors is not an API swap. Anthropic's own tokenizer change makes identical text cost 30% more tokens, and that is the cheap part.
13 min read · Phillip Pham
Claude and Anthropic
Claude EU data residency, endpoint by endpoint: only Bedrock and Vertex AI keep inference in Europe today. What Foundry and the Anthropic API do instead.
13 min read · Phillip Pham · updated 21 Sept 2026
Data foundation
The demo passed on curated files. Permission-aware retrieval fails when the file share has no permission metadata on each chunk. Test that before the next pilot.
12 min read · Noel Dinger · updated 29 Sept 2026
Private AI
What four-bit and three-bit quantisation cost in accuracy, the KV cache and power arithmetic behind the decision, and how offline retrieval stays permission-aware.
10 min read · Noel Dinger
RSS feed · every new post, in any feed reader
Every post hangs off exactly one business area, and each area has a hub that collects everything written about it. The list above shows the same posts by date, so you can see what is new without losing the subject a post belongs to.
That is partly for readers — if you came here because you have to run models inside your own network, you should not have to scroll past six posts about SAP extraction to find the two that matter. It is also the honest structure for the subject. These areas do not overlap much in practice: the people fighting a permission model in SharePoint are rarely the same people negotiating GPU capacity.
What you will not find here: benchmark tables lifted from a vendor deck, customer names we have not been given permission to use, or a post that ends in a demo request. Where we cite a number, the source and its date are on the page.
Field notes on deploying Claude across an organisation: access control, connecting internal systems over MCP, measuring real usage, and reviewing agent-written code.
Building the retrieval layer under a company's AI: RAG architecture, chunking, embeddings, vector stores, evaluation and the rollout that survives a security review.
Posts on the unglamorous half of enterprise AI: extraction from SAP, MES and PLM, masking, entitlements that survive retrieval, and schemas nobody documented.
Field notes on self-hosted inference: picking an open-weight model, sizing GPUs, serving them in production, and what an in-house AI stack costs to operate.
Because a date-only feed buries the useful posts within weeks. Every post belongs to one business area hub, which keeps a subject in a single place for readers and for answer engines resolving a topic to a source. The list on this page adds the date view on top, filterable by area.
The engineers who run the engagements, not a content agency. Posts a named engineer has agreed to sign carry that byline; the rest are published under the company rather than under an invented name. Anything we have not built ourselves is described as an approach, not as experience we already have.
Thirty minutes with the architect who would actually run the engagement.