Five production deployments of an AI agent platform inside customer environments.
Client names are withheld. The architectures, the problems, and what shipped are exactly as they ran.
Anonymized · current as of October 2026 · Greater Phoenix, AZ · Remote-ready
Why this document exists. A forward deployed engineer is judged on whether the thing
works inside someone else's environment, not on whether it compiles. Every case below shipped to a real
customer, ran in production, and was supported afterward. Client identities are removed; nothing else is.
Deployment
Environment
Result
Multi-tenant AI CoS
Own product · SMB clients
Multiple paying clients running concurrently on shared infrastructure with zero cross-tenant context.
White-label partner deployment
Channel partner · reseller model
The partner runs concurrent prospect trials with the upstream vendor appearing nowhere in the customer's view.
Regulated-client deployment
Financial services · client-owned infrastructure
Production CoS running on client-owned infrastructure behind a hard boundary.
Greenfield client build
Professional services · spec to live
Spec to production in days.
Per-member tenant platform
Platform · member-level isolation
Per-member isolation on a shared platform, with tenancy enforced at the runtime boundary instead of relying on the application to behave.
Own product · SMB clients
Multi-tenant AI CoS
Small businesses each need an always-on AI operator, but they cannot share state — one client's credentials, memory and business context can never leak into another's.
What was deployed
One isolated agent profile per client — separate credentials, memory store, skills and job schedule.
Channel wiring across Mattermost, Telegram, Google Chat and iMessage, per client.
Per-client gated demo environments so prospects can evaluate without touching production.
Daily health and keep-alive jobs so no deployment silently goes dark.
Outcome. Multiple paying clients running concurrently on shared infrastructure with zero cross-tenant context. Onboarding is a repeatable sequence rather than a bespoke build each time.
Channel partner · reseller model
White-label partner deployment
A partner resells the service under their own brand. They must never expose the underlying platform vendor to their customers, and prospects need to try it without shell access.
What was deployed
Code-gated trial profiles with a fixed expiry and no terminal access.
All branding stripped to the partner's identity throughout the customer experience.
Per-prospect isolated teams inside the partner's own collaboration server.
Automated gate expiry and teardown so stale trials don't accumulate.
Outcome. The partner runs concurrent prospect trials with the upstream vendor appearing nowhere in the customer's view. White-label held under real commercial load, not in theory.
Financial services · client-owned infrastructure
Regulated-client deployment
A client in a regulated space required their own infrastructure and would not approve shared tenancy under any circumstances.
What was deployed
Separate host provisioned and operated on client-controlled hardware.
Dedicated accelerator, timezone configuration, and its own administrative surface.
Isolated persona defined for that client alone.
No shared state, credentials or memory with any other deployment.
Outcome. Production CoS running on client-owned infrastructure behind a hard boundary. The client retains control of the environment; the compliance decisions stay with them, by design.
Professional services · spec to live
Greenfield client build
A client needed a working system, not a proposal. They also wanted to audit what had actually been installed rather than trust a black box.
What was deployed
Repurposed existing hardware instead of provisioning new.
Full build from written spec through to live: profile, persona, memory seeds, scheduled rituals, channel wiring.
A passcode-gated build-files portal generated directly from the installed bundle, so every file could be inspected.
Operator runbook written alongside the build, not after it.
Outcome. Spec to production in days. The client received a documented, auditable system with a build artifact that matches what is actually installed on disk.
Platform · member-level isolation
Per-member tenant platform
A platform sells seats to individual members, each of whom needs isolated AI assistance — isolation has to be real, not a UI filter.
What was deployed
Tenant-scoped profiles keyed to member identity.
Account plugin handling per-tenant account data.
Isolation enforced at the profile layer where the agent executes, rather than at the application layer.
Outcome. Per-member isolation on a shared platform, with tenancy enforced at the runtime boundary instead of relying on the application to behave.