CASE STUDY · CSI AUTOMATION

M.A.X. — the agentic back office
that runs a one-man operation 24/7.

Before I propose AI for your plant, here's the fleet I trust with my own operation — architecture, discipline, and honest numbers. Names sanitized; patterns real.

The situation

One engineer, no staff, and a day already owned by client work — plus a personal operation with real moving parts: market-risk monitoring against live positions, a research knowledge base, a weekly publishing pipeline, AI/API spend across multiple services, and a website with lead capture. The classic answer is "hire help." The engineering answer: treat the back office as a plant, and automate it with the same discipline used on real steel for 27 years.

The fleet (running today, on one Linux box)

AGENT 01Pre-market risk sentinelEvery morning before the open: pre-market tape vs. every live stop, per-symbol earnings-gap radar, broker-drift checks. Strictly read-only — it recommends, a human acts.
AGENT 02Trade journal engineScores the complete order history into R-multiples, split-adjusted, era-gated so old noise can't pollute current stats. Incremental run: ~6 seconds.
AGENT 03Knowledge-base pipelineNightly ingest of source documents into a queryable engineering KB — with a hard nightly query cap so the bill can't run away.
AGENT 04Publishing automationNightly newsletter/post drafter, retry-hardened after a real API-failure incident; a voice-lint gate keeps machine drift out of the human's voice.
AGENT 05Full-duplex ops channelThe whole fleet reports to a phone via messaging bot; the operator tasks it from anywhere. Failures alert loudly — an outage is a notification, not a silent gap.
AGENT 06Cost sentinel & housekeepingWatches AI/API spend (headline catch: a $500+/mo misconfigured monitoring loop), config-integrity watchdog, secret scanning, weekly log rotation.

The design rules — the part that transfers to your plant

RULE 1Read-only by defaultWrites go through a human hand. The market sentinel has watched thousands of pre-opens and has never placed an order.
RULE 2Hard cost governorsQuery caps and spend sentinels installed before the first surprise bill, not after.
RULE 3Fail loudSilence is the enemy. Every job alerts on failure to a channel a human actually reads.
RULE 4Retry-harden from real incidentsEach production failure becomes a permanent guard — the way a plant turns incidents into interlocks.
RULE 5Small, observable unitsTimer-driven single-purpose agents with logs — not one monolith nobody can debug at 2 a.m.
RULE 6Secrets hygienePre-commit scanning on anything that auto-publishes; credentials never ride along with code.
Design rule I won't break: AI reads everything, recommends often, and writes to production only through a human hand — until months of boring reliability say otherwise.

What this looks like in a manufacturing operation

The same patterns, pointed at plant data: historian and alarm-flood triage briefs before the morning meeting; shift-report and maintenance-log drafting with human sign-off; a maintenance-history mining agent that reads every work order so your planner doesn't have to; energy and utility cost sentinels; an SOP knowledge base with query caps and citations. Read-only first. Human-gated writes. Nowhere near the safety PLC.

See the engagement ladder →