/work · eval results
Each one comes with a full write-up: the architecture, the code and the numbers. Benchmarks are measured, not projected.
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.
Frontier models plan and decompose; a Kubernetes worker pool of local models executes. A shared task ledger, lane-based routing and KEDA autoscaling hold it together.
A containerised pipeline in Dagger's TypeScript SDK that runs identically on a laptop Kind cluster and in GitHub Actions, promoting builds from develop to staging to production on EKS.
A full-stack feature-flag platform with A/B testing and billing, built with Claude Code as a pair programmer: an early field test of how far AI-assisted development goes.
Σ · Tools in the weights