I designed and operate a guardrails-first network of AI agents across six
machines, used daily by non-technical family members. Allowlist/deny
permission architecture, human-approval gates for anything destructive or
outward-facing, an offline-queued message hub, wake-on-mention agents, and
self-built MCP servers -- including a from-scratch OAuth 2.1 remote-MCP
stack (RFC 8414/9728, PKCE).
- 6machines, one governed fleet
- Dailyreal use by non-technical users
- Human-gateddestructive, outward-facing, and spend actions
Read the case study →
Pharmacy-ops master-list optimization
Rebuilt a pharmacy operations quoting workflow around a validated master
list: mapped the process, found where errors entered, and redesigned the
data path so the fast way became the accurate way.
- 5 min → 1 minquote turnaround
- 12% → 1%error rate
- +35%team capacity
food-of-the-week
A live meal-planning product with real weekly users. An agent-in-the-loop
planning flow mines an archive of past weeks, proposes the week's meals,
and a human dials them in -- the agent drafts, the family decides.
- Livein production, every week
- Agent-in-the-loopdrafts proposed, humans approve
food.schuepbach.work
source