Understand
Architecture, research, and technical guides that explain how the stack behaves.
- Architecture explainers
- Original research
- Technology deep dives
AI engineering · Infrastructure · Operations
Understand. Build. Operate.
The AI Stack.
Engineering guides, production infrastructure, hands-on labs, and AI-powered operations tools for the systems behind modern AI.
- Stack layers covered
- 8
- Working tools
- 5
- Resources published
- 23
- Auto-generated pages
- 0
The map
Eight layers, one system
Every layer assumes something about the layer below it, and production is the process of finding out which assumption broke. Select a layer to see what it owns and where it fails.
The AI Stack
8 layers · select one
L8 · AI Engineering
ApplicationsAgents
The product surface. Agent loops, tool calling, and orchestration — where non-determinism meets a user who expects a correct answer.
- Agent orchestration
- Tool calling
- Structured outputs
- Guardrails
The workbench
This is what the work actually looks like
Not a dashboard full of green checkmarks. A real diagnostic path: signals, correlation, a specific hypothesis, and the commands that confirm or kill it.
PRODUCTION INCIDENTWorked examplecloud2 / OpenStack 2025.1
Reported problem
Volume creation fails immediately. Every new volume lands inerror state.
Service health
cinder-api3 hostsHealthy
cinder-schedulerup 4mWarning
cinder-volume@cephlast seen 41mDown
rabbitmqcluster okHealthy
Evidence
$ openstack volume service list
+------------------+---------------+---------+
| Binary | Host | State |
+------------------+---------------+---------+
| cinder-scheduler | ctl1 | up |
| cinder-volume | ctl1@ceph | down |
+------------------+---------------+---------+
$ docker logs --tail 3 cinder_scheduler
WARNING cinder.scheduler.host_manager
No backend capabilities reported for ctl1@ceph
ERROR cinder.scheduler.filter_scheduler
No valid backend was found. Exceeded max scheduling
attempts 3 for volume 8f2c...Probable root cause
The scheduler has no candidate backends. cinder-volume stopped reporting capacity, so capacity filtering returns zero valid candidates and every request fails scheduling rather than provisioning.
The API is healthy and the scheduler is running, which is why dashboards stay green. The failure is a silent one: a service that is up but no longer participating.
What to check, in order
- 1Confirm the volume service is genuinely down rather than disabled
- 2Check the backend connection from the volume host, not the API host
- 3Verify the scheduler is receiving capability reports over the message bus
- 4Confirm capacity filters are not excluding every backend
Next
The analyzer runs entirely in your browser. Pasted output never leaves your machine.
What you can do here
Understand, build, operate, master
Four modes of work, each backed by real material rather than a marketing category.
Build
AI engineering, automation, infrastructure, and deployment practice you can apply directly.
- Reference architectures
- Working examples
- Infrastructure as code
Operate
Incident response, monitoring, runbooks, and the diagnostic method behind them.
- Runbooks
- Diagnostic workflows
- Operations tooling
Master
Academy learning paths, hands-on labs, and validated progression.
- Learning paths
- Scenario labs
- Validation exercises
AI operations workbench
Tools that run in your browser
Paste a Terraform plan, a Dockerfile, or command output and get back an analysis. Nothing you paste leaves your machine, which is also why these work without an account.
AvailableRuns in your browser
Terraform Plan Analyzer
Find destructive and high-risk changes in a Terraform plan before you apply it.
Terraform · OpenTofu
AvailableRuns in your browser
Docker Production Auditor
Check a Dockerfile against production-readiness rules.
Docker · OCI · Containers
AvailableRuns in your browser
OpenStack Health Analyzer
Turn service and agent listings into a ranked view of what is actually broken.
OpenStack · Nova · Neutron
AvailableRuns in your browser
Alert Rule Generator
Generate Prometheus alerting rules that will not page you for nothing.
Prometheus · Alertmanager · Grafana
AvailableRuns in your browser
Prompt Workbench
A searchable library of engineering prompts, kept inside the application.
LLM · Incident Response · Kubernetes
Not yet built
Incident Analyzer, Infrastructure Review, Runbook Assistant — these need server-side inference we have not stood up. They are listed so the roadmap is visible, not because they work.
Inside The AI Stack Academy
One excellent learning path beats twelve shallow ones
Academy tracks are sequenced: objectives, architecture, commands, failure modes, a hands-on exercise, and a validation step you can actually run. We launch a track when it is complete, not when its landing page is.
Learning path
OpenStack Production Engineer
A sequenced learning path for engineers who operate OpenStack in production — architecture, service-by-service depth, deployment, and troubleshooting under pressure.
intermediate· 5 minOpenStackNova
Lab
Lab: CrashLoopBackOff
A service crash-looping after a routine config change, with logs that show a normal startup and no error anywhere. Work out what is killing it.
intermediate· 25 minKubernetes
Lab
Lab: Nova NoValidHost
Instances stopped building on a cloud with visible free capacity. Work the evidence, form hypotheses, and find why the scheduler has no candidates.
expert· 35 minOpenStackNova
Operate
Procedures and practice
Runbooks written to be followed under pressure, and labs that drop you into the failure before telling you the answer.
Production runbooks
All runbooks →Runbook
Cinder — No Valid Backend
Diagnose and remediate OpenStack volume creation failures caused by the scheduler having no candidate backends, including the case where the API reports healthy.
expert· 20 minOpenStackCinder
Runbook
CrashLoopBackOff Triage
A structured procedure for a pod that will not stay running, covering exit-code classification, probe-induced restarts, OOM kills, and validation that the fix held.
intermediate· 15 minKubernetes
Runbook
Linux Disk Pressure Response
Find what is consuming a full filesystem, reclaim space safely including the deleted-but-open-file case, and validate before the host causes wider failures.
intermediate· 15 minLinuxKubernetes
Hands-on labs
All labs →Lab
Lab: CrashLoopBackOff
A service crash-looping after a routine config change, with logs that show a normal startup and no error anywhere. Work out what is killing it.
intermediate· 25 minKubernetes
Lab
Lab: Nova NoValidHost
Instances stopped building on a cloud with visible free capacity. Work the evidence, form hypotheses, and find why the scheduler has no candidates.
expert· 35 minOpenStackNova
Lab
Lab: Destructive Plan Review
A routine-looking Terraform pull request. Find the changes that would destroy data before you approve it, and work out which one is not what it appears to be.
intermediate· 20 minTerraformAWS
Cornerstone resources
Start with these
Deep, consolidated guides rather than a page per keyword. Each one is intended to be the last thing you need to read on its topic.
Guide
Docker Production Engineering
Building container images for production: reproducible builds, correct signal handling, non-root runtime, layer strategy, and keeping secrets out of image history.
intermediate· 4 minDockerContainers
Guide
GPU Infrastructure for AI
What determines accelerator performance in production: memory capacity versus bandwidth, interconnect topology, and the checks that find a misplaced workload.
advanced· 6 minGPUNVIDIA
Guide
OpenStack Production Operations
Operating OpenStack in production: service state versus status, the message bus, placement disagreements, and the quiet failures that keep dashboards green.
expert· 3 minOpenStackNova
AI infrastructure
The layer with the least honest writing
Accelerators, interconnects, storage, and the private cloud control planes underneath them — written from the operator's side.
Guide
AI Networking Fundamentals
Collective communication, RDMA, and rail topology explained from the operator's side, including why adding nodes to a training job can make it slower.
advanced· 4 minRDMAInfiniBand
Guide
GPU Infrastructure for AI
What determines accelerator performance in production: memory capacity versus bandwidth, interconnect topology, and the checks that find a misplaced workload.
advanced· 6 minGPUNVIDIA
Guide
Kolla-Ansible Production Architecture
How Kolla-Ansible actually structures an OpenStack deployment, where configuration comes from, and how to make changes without discovering them during an outage.
expert· 3 minOpenStackKolla-Ansible
Research
Measured, not asserted
Research is published with its methodology, its environment, and its raw data. Projects that have not been run yet are listed as exactly that — announced, not pretended.
Research
AI Infrastructure Operations Survey
A planned survey of how teams actually operate AI infrastructure in production — what breaks, what is measured, and where engineering time goes. Not yet conducted.
foundational· 2 minAI InfrastructureGPU
Research
GPU Cloud Price Index
A planned recurring measurement of accelerator pricing across cloud providers, normalised so the comparison is meaningful. Methodology published before data collection begins.
foundational· 2 minGPUCloud
Why trust this
Technical proof instead of testimonials
We have no customer logos to show you and will not invent any. What we can show is how we work.
Every claim is attributable
Author, reviewer, published date, updated date, and last-verified date on every technical resource. Where a procedure was executed, the exact versions are listed. Where it was not, we say so on the page.
How we test →Pages earn their place in the index
A URL becomes indexable only after passing a ten-point editorial gate. Most of what we hold stays inside the application, searchable here and absent from the sitemap. Database size is not sitemap size.
Editorial policy →AI assists; humans are accountable
AI helps with research organisation, drafting, and editing. No technical content is published as reviewed without a human engineer reviewing it, and pages awaiting review say so in a banner.
AI use policy →Newsletter
Inside The AI Stack Brief
A practical weekly briefing on AI engineering, infrastructure, production operations, and the technologies powering the AI stack.
One email a week. No sponsorship placements inside the technical sections. Unsubscribe in one click.
Pro
Full Academy, advanced labs, saved work
Everything in the Academy, advanced labs, the full runbook library, and saved investigations once accounts ship. Free access covers the guides, the working tools, and the introductory labs.