The $6M Lambda Swarm — Step Functions Hierarchy + S3 JSON State
┌─ Master Step Function
│ pulls work · fans out · monitors
│
├──→ Worker SFN ──→ Lambda × N
├──→ Worker SFN ──→ Lambda × N
└──→ Worker SFN ──→ Lambda × N
│
└─ S3 JSON state (recovery primitive · resume-from-anywhere)
August 2024. Streamlit Revenue Opportunity engine on a 3-hour feedback loop with a US-based CEO and $6M on the line. Hierarchical Step Functions over flat fan-out; S3 JSON as the cheapest possible state store; first meaningful Claude collaboration on architecture. 2hr batch → 5min. The round closed.
2025. The Win Agent shipping at $3.5M ARR scale, with "change the prompt and pray" as the inherited test discipline. Replaced it with invariant-based validators + drift sentinels + replay cache. Same Cynical Architect doctrine, applied to the LLM layer. ADR-0013 + ADR-0015.
Terraform consolidation ──→ clean diff · state-files committed
│
▼
CI/CD pipelines stall · 1hr+ war room
│
──── the seam ──── legacy env × new NestJS/NX
│
5 minutes: pattern-match · fix · pipeline up
Late 2025. The Infra-Flip shipped clean — the framework view said the system should run; the system did not run. The fix lived at the seam between the legacy environment and the new conventions. ADR-0016 named the doctrine the moment taught.
One index.html. One stylesheet. One ~120-line vanilla-JS markdown parser. Zero build step. Zero node_modules. Zero dependencies.
18 ADRs at /adr explain every structural decision. ADR-0001 is the meta-decision; the rest follow.
The medium is part of the message. If a hand-rolled monospace site reads as a trust signal, this audience reads me right.
Six-year volume closed May 2026 on the company's own architectural floor — Scheduling & Payments v2 shipped 30 days before exit. The exit itself was an act of integrity, not a career move; the doctrines stay, the work compounds. Phase 11 →
Available Jun 2026. Open to Staff Engineer, Tech Lead, Engineering Manager, or Founding Engineer roles where the work crosses system seams.