Das FDE-Deployment-Toolkit 2026
Scope, cost and go-live-ready a production AI deployment in days — without rebuilding the four documents you rewrite on every single deployment.
Referenzarchitekturen, eine 60-Punkte-Go-live-Checkliste, GPU-Kostenmodelle und eine Anbieter-Shortlist — die vier Artefakte, die du bei jedem Deployment neu baust.
Indicative price only — not a final sale price. Checkout is not open yet; the final price requires owner confirmation before anything can be bought.
- 12 Referenzarchitekturen (On-Prem, VPC, Edge)
- 60-Punkte-Go-live-Checkliste
- GPU-Kostenmodell als Tabelle
- Anbieter-Shortlist mit Abwägungen
Bestellung, Zahlung und Dateizustellung sind gebaut und im Testmodus Ende-zu-Ende geprüft. Echte Käufe bleiben deaktiviert, bis der Inhaber das Zahlungskonto verifiziert und die Endpreise freigibt. Der genannte Preis ist nur ein Richtwert, kein freigegebener Verkaufspreis.
Delivery system status & test record →Für wen — und für wen nicht
For
- Forward-deployed, deployment and solutions engineers running AI deployments
- Technical leads and architects who need a defensible go-live checklist with evidence and owners
- Anyone weighing GPU vs hosted-API cost who needs a model with real inputs, not a vendor pitch
Not for
- Not a course or video — it is documents and a spreadsheet you fill in yourself
- Not vendor-specific: there is no tie-in to one cloud or one model provider
Was du wirklich bekommst
toolkit.md≈ 72 KBMain document, four parts: 12 reference architectures (with a combination matrix) / a 60-point go-live checklist (with evidence and owner columns) / the GPU cost model / a vendor shortlist with a negotiation question set; plus a discovery-phase question bank, a capacity-calculation sheet, and a glossary. It contains no prices, benchmarks, or case figures anywhere.toolkit.html≈ 96 KBA4 print edition generated by md2html; export to PDF via Ctrl/Cmd+P in the browser (enable background graphics).gpu-cost-model.csv≈ 4 KBThe cost model itself: 16 input rows (all PLACEHOLDER) + 18 formula rows; change an input in Excel / Google Sheets and every formula recomputes.
Listed straight from the shipped toolkit.md + manifest.json — nothing below is a mock.
Inhaltsverzeichnis
- What is inside
- How to run an engagement with this toolkit
- Reading conventions
- Part 1 — Twelve reference architectures
- 1.0 How to choose one
- 1.1 A1 — Managed single-VPC baseline
- 1.2 A2 — Multi-account landing zone with a central egress hub
- 1.3 A3 — Private connectivity to the model provider
- 1.4 A4 — Hybrid: on-prem data plane, cloud control plane
- 1.5 A5 — Air-gapped on-prem
- 1.6 A6 — Cloud-burst GPU behind an on-prem backplane
- 1.7 A7 — Edge inference with a local retrieval cache
- 1.8 A8 — Regional split by tenant
- 1.9 A9 — Multi-tenant isolation in one shared deployment
- 1.10 A10 — Model gateway and routing layer
- 1.11 A11 — Retrieval overlay on the existing system of record
- 1.12 A12 — Agent sandbox with a hard side-effect boundary
- 1.13 Comparison matrix
- Part 2 — The 60-point go-live checklist, with evidence
- Architecture and boundaries
- Data and retrieval
- Evaluation
- Observability
- Security and access
- Cost control
- Go-live process
- Rollback and incidents
- Part 3 — GPU cost model
- 3.1 Why GPU cost models fail
- 3.2 Units and vocabulary
- 3.3 The model
- 3.4 The inputs, and where the truth actually lives
- 3.5 The spreadsheet
- 3.6 Worked example
- 3.7 Break-even: hosted API or your own hardware
- 3.8 Sensitivity: which lever to argue about
- 3.9 Cost per resolved task
- 3.10 Twelve cost traps
- Part 4 — Vendor shortlist with trade-offs
- 4.1 What actually decides the choice
- 4.2 The categories, and who fits them
- 4.3 Trade-off dimensions, side by side
- 4.4 Fourteen questions to ask in procurement
- 4.5 Shortlist worksheet
- 4.6 Keeping this list honest
- 4.7 What this part deliberately does not claim
- Appendix A — Discovery question bank
- Appendix B — Sizing worksheet
- Appendix C — Glossary
- Appendix D — Changelog
- Appendix E — Licence
Ein echter Auszug
What is inside
· Part · Document · Use it when · · 1 · 12 reference architectures · Week 0. Choosing the deployment shape before any code exists · · 2 · The 60-point go-live checklist (with evidence and owners) · Week 2 onward. The gate between "works in the demo tenant" and "runs for real users" · · 3 · GPU cost model + spreadsheet · Week 0 for the business case, then monthly · · 4 · Vendor shortlist with trade-offs · Week 0 and every renewal · Appendices carry the discovery question…
How to run an engagement with this toolkit
· Phase · What you actually do · Use · · Week 0 — discovery · Interview the business owner, read the environment, pick a deployment shape, produce the first cost model · Part 1, Part 4, Appendix A · · Week 1 — thin slice · One workflow, one team, real data, deliberately narrow. The goal is to find the failure modes while they are still cheap · Part 1 (the chosen architecture's gate list) · · Week 2 — harden · Evaluation set frozen, observability in place, permissions…
Excerpted verbatim from the actual toolkit.md — not sample/mock content.
So nutzt du es
- Open toolkit.md in any Markdown viewer, or print toolkit.html to PDF (browser Ctrl/Cmd+P, enable background graphics).
- gpu-cost-model.csv opens in Excel or Google Sheets — change the 16 input cells and the 18 formula rows recompute on their own.
- Every cost input is a placeholder on purpose. Fill it from your own pricing — no vendor prices, benchmarks or case figures are baked in.
Version & Updates
Version 1.0 · released 2026-09-27 — Version 1.0, released 2026-09-27. When we correct an error or add material, we bump the version and date the change — never a silent rewrite.
Lizenz
Personal and internal team use. You may use and adapt these materials for your own deployments. You may not resell, redistribute or republish the toolkit, in whole or in part, as a standalone product.
Häufige Fragen
Do I need a specific cloud or vendor?
No. The reference architectures cover on-prem, VPC, hybrid and edge; the cost model is vendor-neutral — you enter your own prices.
Does it contain code?
The cost model is a spreadsheet and there is a worked cost example. It is not a code library — it is the documents around the build.
Is it suitable for beginners?
It assumes you can deploy software and read an architecture. It will not teach you what a VPC is.
Is it still updated?
Yes. The changelog carries the current version and dates. Corrections are versioned, never silent.
RAG in Produktion — Playbook
Chunking-Strategien, Reranking-Setups, Evaluations-Harnesses und die sieben Fehlerbilder, an denen jede RAG-Demo in Woche drei stirbt.
30 KI-Deployment-Muster aus aktuellen FDE-Stellenausschreibungen
Dreißig Deployment-Muster aus echten, aktuellen FDE-Stellenausschreibungen — jede mit dem ursprünglichen ATS-Link der Arbeitgeber, einer redaktionellen Einordnung und einer klaren Liste dessen, was diese Daten nicht belegen.
Enterprise-Prompt-Paket — 208 aufgabenspezifische Vorlagen
208 originale, aufgabenspezifische Prompt-Vorlagen für acht Enterprise-Workflows, jede mit Variablen, einem expliziten Ausgabevertrag und einer ehrlichen Angabe dessen, was nicht benchmarkt wurde.