What we build
Production AI capabilities, not demo theater.
Orchestration, evals, guardrails, cost awareness, and systems that survive operators — backed by shipped work in The Lab.
Agent orchestration
Supervisor patterns, tool-calling loops, failure recovery, and knowing when a pipeline beats an agent.
Proof: Estate Mogul OS routing lanes before LLM fallback
MCP & tool integration
Connectors for workspace, CRM, email, and webhooks with isolated token stores and explicit scopes.
Proof: Google Workspace OAuth separated from core runtime secrets
Eval & verification
Regression tests, smoke gates, and structured checks before production deploy — not vibe-based QA.
Proof: validate_build.sh + pytest suites on Travis runtime
Guardrails & agency limits
Preview-confirm flows, connector honesty, least-privilege tools, and kill-safe outbound actions.
Proof: Gmail send and sheet mutation require explicit confirmation
Production observability
Health endpoints, log scans, runtime verifiers, and operator-visible job states.
Proof: Knox /api/health and HotlistVerify queue visibility
Cost-aware model routing
Route by task complexity, cache where it matters, and design loops with inference economics in mind.
Proof: LiteLLM routing layer on self-hosted VPS stack
Full-stack product delivery
Next.js apps, APIs, auth flows, admin tools, email, and deployment — end to end.
Proof: Knox Client Portal and multiple live client domains
VPS & Docker operations
systemd, nginx, backups, compose profiles, and rollback-friendly deploy habits on Linux.
Proof: Production services on owned VPS infrastructure
Workflow automation
n8n scaffolds, Zapier edges, cron sync, and staged activation instead of reckless auto-run.
Proof: Disabled-by-default n8n imports with validation runbooks
Human handoff design
Automation for busy work, explicit moments where a person must step in.
Proof: Preview-confirm queues and connector honesty messaging
See it in shipped work
Every capability maps to a case study with real constraints.