Long-form engineering notes on self-hosted models, agent fleets, GPU platforms and the economics of inference, written by someone who runs them. Part blog, part portfolio, part model card.
Type help or try
01 / Latent space
30 posts embedded, projected and clustered. Hover a point to see its neighbours, filter by topic, or turn up the temperature in the header and watch it drift.
↵ read
Each point is a post. Position comes from a build-time embedding of its tags and title, pulled together by cosine similarity. Size is length in tokens. Lines join nearest neighbours.
02 / Recent outputs
03 / Long context
8 parts · 47,836 tokens
From an empty VPC to GPU node pools, vLLM, KEDA and a cloud-vs-local inference router, built as reproducible AWS CDK you can deploy and tear down yourself.
EKS · AWS CDK · vLLM · KEDA · Karpenter · Hono
8 parts · 26,455 tokens
Turn one Firecracker host into a fleet: a control plane that places agents onto bare-metal AWS hosts on demand, and a routing layer that gets requests back to them.
Firecracker · AWS CDK · EC2 Metal · Lambda · DynamoDB
04 / Evals
A control plane that places eve.dev agents into Firecracker microVMs across a fleet of bare-metal AWS hosts, plus the routing layer, snapshot-based warm starts, a deploy CLI with per-PR preview agents, and the FinOps view of cost per agent.
VPC to GPU node pools to a TypeScript router that sends each request to local vLLM or to Claude, with KEDA scaling GPUs to zero overnight and Langfuse showing the dollar cost of every call.
Document Q&A with citations, invoice reading, agent-drafted quotes with human approval, and a shop assistant, all on Workers AI, Durable Objects, Workflows and D1, each with a per-request cost ledger and a budget in front of every model call.
Retries with backoff, a circuit breaker per destination, duplicates delivered once and a dead-letter list, written in Effect and run on Workers, Queues and Durable Objects, with infrastructure defined in the same program via Alchemy.
An auth and rate-limit gateway, gVisor sandboxing for agent tools, prompt-injection guardrails, scoped secrets and signed model weights, mapped to the OWASP LLM Top 10.
05 / Next token