AI engineering · a one-person company
Learning.
I’m one person. CommsCrew is what happens when you take that seriously — a corporate comms department that’s itself a crew of AI agents, built and shipped by AI agents, run end-to-end by a single operator. These are the AI-engineering decisions behind it: how each part works, what it cost, and what running a one-person, agent-operated company actually taught me.
Before you read
These are my own notes and opinions — not my employer’s. The whole point of this build was to see how far one person could get on a near-zero budget, so every decision optimizes for cheap and educational, not for enterprise scale, uptime guarantees, or compliance. Where the enterprise answer would be different, I try to say so. Copy the reasoning here, not the config.
The thesis
How much of a company can one person run if agents do the labour? Agents show up here at two layers. Inside the product, four of them — Strategist, Writer, Editor, Analyst — do the work of a comms team. Outside it, the build and the deploy were driven by agents too; I was the operator, not the typist. The five sections below are the AI-engineering decisions that made both layers hold together on a near-zero budget — grounded in the real code, honest about the failure modes.
Where this is going
The two layers above are what runs today. The direction is a third: a governed fleet of agents that handle the business functions themselves — a control-plane watching cost, evals, and access; an engineering relay that takes an issue from triage to release; an operating loop from founder vision to shipped feature. That work is in progress and built in the open, tier by tier — so I’m not going to claim it’s running here until the receipts say it is. When it earns its own writeup, it’ll be badged for what’s live versus reference.
System architecture
Read →Monorepo to request lifecycle to SSE to memory-in-Postgres — and how dev and prod stay the same shape.
“SSM tag pointers decouple what's deployed from what CDK last synthesized.”
The multi-agent crew
Read →Four agents, a deterministic orchestrator, memory injection, the prompt-cache layout, and the two output contracts.
“Specialized prompts with structured handoffs through a deterministic coordinator — chosen for debuggability and cost control.”
The backend
Read →One async codebase across SQLite and Postgres, UUID keys, JWT, and the edit-signal → brand-voice learning loop.
“Match output-format rigor to the consequence of a bad parse, not uniformly.”
The frontend
Read →Route groups, token-by-token SSE consumption, and the design-token system that themes every surface.
“Robust SSE parsing is about the boundary between network chunks, not the protocol.”
Deployment & the $10 stack
Read →The cheap stack, keyless OIDC, the migrate-gated alt-port swap, and auto-rollback — near-zero-downtime on one box.
“SSM is the source of truth — so the deploy path is the disaster-recovery path.”