Exploring the intersection of AI, SAP, and enterprise automation through research-driven perspectives and practical experience.
At the DSAG-Jahreskongress in Cologne, the German-speaking SAP user group asked SAP not to tie AI to cloud contracts. I checked SAP's own documentation, its announcements and the SAP PRESS handbook to see what on-premise S/4HANA can use today: some AI runs, some is gated behind a cloud commitment, and some is missing.
At the DSAG Jahreskongress in Cologne, about one in three items in the program references AI, and SAP has built agents deep into its procurement products. A new preprint on management illusions explains why many programs still stall, and how sponsors can tell good use cases from bad ones.
PageIndex promises retrieval without embeddings: index a document as a tree of its own sections and let an LLM reason its way to the right one. I measured it against six conventional pipelines on the same corpus, the same questions and the same judge. It lost, and the way it lost is more interesting than the fact that it did.
My local model stopped producing language and emitted only exclamation marks. I traced it to NaN logits, blamed speculative decoding, ran 150 controlled samples, and wrote it up as a dead end. Then I pulled a newer upstream image and found a code comment warning against the exact one-line patch I had written myself.
My Literature Share on the Synology holds hundreds of documents and answers nothing. I rebuilt retrieval around a reasoning-based idea instead of a vector index: no embeddings, no chunking, one model endpoint on the box. And I checked the benchmark everyone quotes, because the table behind it says something the headline does not.
Most RAG pipelines are blind to charts, diagrams and screenshots, which is exactly where business documents carry their answers. How a bar chart actually gets into a vector space: Docling for parsing, then three routes with very different cost profiles.
I read the SAP PRESS handbook Künstliche Intelligenz in SAP Systemen cover to cover. It answers the question every SAP estate eventually faces: some corners look old, but that is not death. Artificial intelligence does not replace the system of record. It runs through it, one reviewed decision at a time.
GLM-5.3-Flash is a 320B Mixture-of-Experts model quantized to NVFP4 — 598.5 GiB down to 181 GiB. It runs on two DGX Sparks under my desk, with correct reasoning and working tool calls. But the measurements tell a narrower story than the spec sheet: concurrency buys literally nothing, prefill is eleven minutes at full context, and speculative decoding fails at weight loading. LibertAI published no throughput numbers, so these are the first measured for this checkpoint.
Qwen3.8-Flash-Next is a 125B Mixture-of-Experts model with an unusual extra: a 51-billion-parameter n-gram embedding table. That table is why vLLM could not serve it on my two DGX Sparks, no matter which of eighteen community quantizations I tried. The fix was a one-line patch, and the road there produced a benchmark result I did not expect: two optimizations that each help, and cancel each other when combined.
I built a macOS menu-bar app to watch my two DGX Sparks. It has exactly one user, and that is why it taught me more about observability than most production dashboards: nothing was hidden behind a vendor abstraction. Six numbers on that panel were wrong, and every one of them was wrong for a reason that applies directly to monitoring AI systems in an enterprise. A CPU metric that confidently lied, a read-only connection that silently corrupted its own reading, and four more.
I put fourteen quantized LLMs on the grid of my two-node DGX Spark cluster and timed every one, top speed measured in tokens per second. A 35B model took pole at 57.1 tok/s, but the real finding sits behind it: a 122B model lands within 29 percent of that, and speculative decoding nearly doubled a 120B into the same class. Total size predicts very little. Here is the full grid, the configuration each car ran, and what actually moves the needle.
A German consortium just published Soofi S, a 30B open foundation model you are actually allowed to inspect: weights, training recipe, and per-source data accounting all on the table. I read the paper. Here is what it can do, why a sovereign model built in Germany matters, and where it earns its place in a business stack.
I run two coding agents side by side and kept re-typing the same corrections. So I read the standards, mined 2,500+ of my own prompts, and rebuilt my setup around one rule: put every instruction where it actually loads. Here is what the evidence says — and what I changed.
Autoregressive decoding hits the same wall a rocket does near light speed: more raw compute buys less and less. DFlash, a new ICML 2026 paper from UC San Diego's Z Lab (arXiv:2602.06036), sidesteps it with block-diffusion drafting instead of brute force. I read the paper, checked the benchmarks, and tested what it actually takes to run it on my own DeepSeek V4 Flash box.
Data Science Dojo and SambaNova ran a hands-on webinar series that builds a production-grade AI agent from a bare ReAct loop all the way to evaluation and deployment. I walked through every notebook — here is the distilled arc: think, build, remember, capable, scale, ship. Code and notebooks are public on GitHub.
Dozens of agentic AI frameworks now compete for attention, and most comparisons mislead — they pit tools from different layers against each other. Here is a clear map: what LangChain, MLflow, Haystack and the leading alternatives actually do, the six dimensions that separate them, and how to choose without locking yourself in.
A RAG pipeline produces fluent answers from day one — whether or not they are correct. Here is how to measure RAG performance properly: retriever, generator and end-to-end metrics, why overlap scores like BLEU mislead, how to wire it all into a test suite with DeepEval, and what a measured benchmark of six RAG pipelines reveals.
Large language models are fluent but not factual. Here is how an RDF triple store like Oxigraph turns a knowledge graph into a verifiable source of truth an LLM can query — and why that beats trusting the model's memory.
Exploring the intersection of Large Language Models and Enterprise Resource Planning — how AI can understand and automate complex SAP workflows.
Building production-ready Retrieval-Augmented Generation pipelines for enterprise environments — from vector embeddings to evaluation frameworks.
How autonomous AI agents are transforming procurement — from supplier evaluation to requisition routing and approval orchestration.
Follow my research and insights on AI-driven enterprise automation