The caskeycoding.com tech stack at a glance
This site is a Next.js static export on CloudFront, a small Python service on Lambda, a single DynamoDB table, an S3 content bucket, and two AI agents calling Anthropic directly. Everything is wired up with AWS CDK across four stacks split by deploy cadence — auth, agents, backend, frontend.
Here is the stack, layer by layer.
Frontend
- Next.js with static export, TypeScript,
trailingSlashaligned with the edge rewrite - Playwright for end-to-end tests
- Lighthouse CI for performance budgets
Edge / CDN
- CloudFront distribution fronting a private S3 site bucket via origin access, HTTPS-only
- CloudFront Functions at viewer-request for extensionless URI rewrites and one retired-slug 301
- Route53 hosted zone with apex /
wwwaliases plus Google Workspace MX, SPF, and DMARC (p=quarantine) - ACM for the TLS certificate
API & compute
- Amazon API Gateway for the dynamic surface
- AWS Lambda (Python) for the blog handler, the public demo handlers, the agent API, and the long-running orchestrator
- AWS WAF with regional, per-route rate rules on the unauthenticated
/public/*endpoints
Data
- Amazon DynamoDB — single-table design,
postId+typekeys,postandagent_taskitems - Amazon S3 — transparent content offload at 2KB so DynamoDB items stay small and read-cheap
Auth
- Amazon Cognito user pool, shared via cross-stack reference to the backend
AI / LLM
- Anthropic API (direct, official Python SDK) as the primary path
- Sonnet 4.6 — workhorse: generation, polish, tool loops
- Opus 4.7 — synthesis and multi-source reasoning
- Haiku 4.5 — routing, classification, eval-judge
- Amazon Bedrock — fallback only, via a cross-region inference profile, with a Discord webhook alert when it engages
- Shared client (
anthropic_client.py) owns retries, fallback, secrets, and cost accounting; agents stay thin
Secrets & config
- AWS Secrets Manager for the Anthropic API key — no env-var secrets in prod
- AWS Systems Manager / CDK context for non-secret config
Observability
- Amazon CloudWatch logs, metrics, and alarms (5xx on the static path → SNS topic)
- Amazon SNS for paging
- Discord webhooks for human-in-the-loop alerts (LLM fallback, eval drift)
- Structured
llm_callJSON logs — full payloads only in non-prod; prod carries metadata plus aprompt_sha256
Testing & CI
- pytest for the Python service
- Replay / eval harness — YAML cases with cached completions, runs on every PR with no API spend; a
--livemode for capturing new fixtures - CI pricing check — fails the build if the in-repo per-model pricing table is older than 90 days
Infrastructure as code
- AWS CDK (TypeScript), four stacks split by blast radius:
AuthStack— CognitoAgentStack— agent and orchestrator Lambdas, Bedrock IAMBackendStack— API Gateway, blog Lambda, DynamoDB, S3 content bucketFrontendStack— CloudFront, site bucket, Route53, ACM
The specs that drive this site — including the architecture decisions behind each of these choices — are public in the specs demo repo.
Keep reading
Welcome: Building Platforms for Scale
An introduction to the blog, reflections on infrastructure monitoring, platform leadership, and building systems that empower organizations to innovate safely at scale.
The Orange Pi That Maintains Itself
A small ARM box that started as a local LLM experiment and ended up a self-governing node: private retrieval, a resident agent under a written constitution, a code-enforced safety fence, and a nightly job where it audits itself and files its own backlog.
An orchestration mode is only as good as its backlog
Anthropic published a guide on building a session-level orchestration mode. I built it two ways, on the CLI and on the API, and then hit the part the guide does not cover: an orchestrator that fans out is useless without a backlog of real work to fan out over.
Wiring Garmin Into My Marathon Coach: A Live Data Integration Without an Official API
How I replaced manual CSV exports with a live Garmin data feed for my AI marathon coach: a scheduled unofficial-API poller, resilient session handling, and the design calls that keep training and recovery data fresh and trustworthy.
A Boring Design Let Me Run a Black Swan on a Tuesday
Two posts ago I bet that keeping my portfolio reviewer's engine deterministic and auditable was worth it. This is where that bet paid off: because the engine is replayable, I could run a simulated market crash through the real production code and catch a money-losing flaw on paper — before it could ever cost a real dollar.
Building a Personal Finance Reviewer: What Survived the Rewrite
A personal portfolio reviewer where the scoring is deterministic and the AI only narrates. The architecture that held up after I had to rewrite the model it was built on, and why that boundary is the whole point.