Final phase of my AWS RAG series: does Bedrock Guardrails' grounding check catch a real citation bug I found by hand in Phase 2? Sometimes — 0.92 uncaught, 0.68 caught. @TSBCanada https://t.co/Ie803z5S7x
Phase 5 of my AWS RAG series: I ran the validation I promised in Phase 4, and it corrected my own number — 39% → 23.6%. @TSBCanada https://t.co/astSKE2BQD
Phase 4 of my AWS RAG series: the "so what" article. 4 findings, 5 actions, each traced to an exact count or TSB quote. Headline number self-corrects on purpose: 39% (curated sample) validated down to 23.6% via a 313-report random sample, for $0.1568.
https://t.co/CcEzpbuXhw
Phase 3 of my AWS RAG series: from retrieval to a tool-calling agent, built on Bedrock Converse's native tool use. Nova Lite routes itself between an exact-count tool and the citation-gated search tool from Phase 2.
https://t.co/WZQ77Vxr9N
Ingestion failed on 22 of 24 documents. Looked like throttling — wasn't. A reserved metadata key needed to be marked non-filterable at index creation, or writes silently fail.
Part 2 of building a RAG pipeline on AWS Bedrock.
https://t.co/Ml6LV67L0f
Building a RAG pipeline on AWS Bedrock over Transportation Safety Board of Canada investigation reports.
Assumed the data was openly licensed. It wasn't — TSB uses Crown copyright, not the standard open license.
Part 1 of 6, documenting real bugs as I go.
https://t.co/D4pLFqt0ns
Closed the loop on my 8-part EKS platform series: AWS Load Balancer Controller turns Ingress objects into real ALBs — no more kubectl port-forward. Build → EKS → GitOps → metrics → autoscaling → public internet. No human past git push. #Kubernetes
https://t.co/rYOj8ecu28
An HPA scales the app on its own requests-per-second (via prometheus-adapter, not CPU), and Karpenter watches for pods that don't fit and launches right-sized EC2 capacity to fit them. Load-tested it live: REPLICAS walked 1 → 25 in ~90 seconds.
https://t.co/bDJGHgVmbZ
Part 6 of my platform-layer series: turning Stage 5's Prometheus metrics into a dashboard nobody has to build by hand.
The dashboard is JSON, checked into git as a Helm chart. ArgoCD syncs it as a ConfigMap, a sidecar container. https://t.co/uQYKkZvSoG
Part 5 of my platform-layer series: making the app's own Prometheus instrumentation — written in Stage 1, unused ever since — actually mean something.
The proof: generate 8 requests
https://t.co/vqOzFz6NkL
Part 4 of my platform-layer series: replacing helm install with GitOps — ArgoCD watching this repo and reconciling a real EKS cluster to match it.
The interesting proof isn't the sync, it's self-heal.
https://t.co/0hgbXbZCBi
Part 3 of my platform-layer series: packaging a FastAPI app + an in-cluster Postgres with Helm on real EKS.
The theme is composing a chart you didn't write — Postgres comes from the upstream Bitnami chart with a values overlay.
https://t.co/TOU6kxP8Jw
Part 2 of building a Kubernetes platform on AWS, on real infra: EKS with Terraform.
Includes the two failures that cost me a terraform apply — node-group AMIs vanishing for aged-out k8s versions, and free-tier instance-type lockouts on new accounts.
https://t.co/0b6SUBAnfE
New series: building a production EKS platform end to end — Terraform, GitOps with ArgoCD, Prometheus/Grafana, Karpenter, ingress — with every real AWS failure documented, not just the happy path.
Part 1: containerizing the app. Non-root, instrumented
https://t.co/QFjkH3nVIs
Save up to 75% on Generative AI Document Summary Solutions by pairing AWS Textract with Claude Opus 4
Significant Savings with respect to your AI Driven Document Management Solutions
https://t.co/1qT75Jjzx1
@Bounteous@AccoliteDigital@awscloud@AnthropicAI
Anthropic's Claude Opus 4 performs very well in Document Analysis!
Please read on if you wish to learn more about this is done with Anthropic's latest Claude release.
https://t.co/9EyNQOYrRo
@awscloud@AnthropicAI@AccoliteDigital@Bounteous