I'm Joe, from Scotland, living in Vancouver.
Monorepo enthusiast. Caremad about developer experience.
4 nodes · 52 CPUs · 112 GB · one RTX 4090
A D&D campaign manager built on grants: the same monster renders full stats for the DM, redacted stats for a player who has fought it, and just a name for one who has only heard rumours. play →
Boot a workload once, freeze it, restore the snapshot for every request. The guest never holds a real secret; an egress proxy swaps placeholder tokens for credentials at the network hop. EmberVM runs this substrate today: one-shot tasks, stateful sessions, and warm HTTP serving. Watch a microVM restore from disk in 22ms. watch it restore →
Tag @Bosun in a Discord thread and a goose agent wakes in its own microVM, runs a recipe, and replies with an artifact or a PR. Ambient chat, scheduled routines, and coding agents all dispatch through fc-invoke (~140ms to first model call). Its successor, EmberVM, already runs the semgrep scans and warm HTTP serving; fc-invoke is frozen and the goose agent is its last tenant.
vLLM serving a 35B sparse-MoE (~3B active), int4-mixed weights with an fp8 KV-cache, ~170 tok/s single-stream decode. Chat, the agents, and the knowledge graph's RAG all share the 4090.
One Postgres backs every app; pgvector indexes the embeddings for a fileless knowledge graph. notes is a public RAG over it. Declarative migrations applied by an operator, volumes replicated across nodes.
Five custom Bazel rulesets build every image dual-arch and pin digests into versioned OCI Helm charts; ArgoCD reconciles the cluster from the repo.