01
Voice & comms agents
Agents that answer calls, triage email, and handle inbound messages end-to-end. They book, follow up, and escalate to a human only when it matters.
twilio · livekit · openai · webhooks
Production agentic AI systems
Agentswork.
Youorchestrate.
Your inbox, handled. Your calls, answered. Workflows running in the background — reachable by voice, text, or API. I build the ones that actually run in production, not just in a demo.
agent · live
vecera/orchestrator10:42:01 [call] inbound +1 415 ••• 2207 → routing
agents 12 runninguptime 99.98%last action 3s ago
systems operationalp95 latency 412msevents/24h 18,304v1.0 — built quietly
01 /
Unattended by default
Systems built to run with no human in the loop — and the observability to prove they're running.
02 /
Reachable everywhere
Voice, SMS, email, webhook, API. One agent, every surface your users actually touch.
03 /
Boring reliability
Retries, idempotency, queues, budgets. The unglamorous engineering that keeps things alive at 3 a.m.
The whole point
Go take the hike.
They've got the desk.
You didn't build this to spend the day answering the same call. Agents take the desk so you can take the hike — catch the kid's game, sleep through the night, run the business instead of living inside it. You orchestrate. The system runs.
That only works if it actually works. Which is the other half of this site — the engineering underneath.
06:14 you’re on the trail
06:14 agent.answers(call #1042)
06:15 agent.books(thu 2:30pm)
07:02 agent.triages(inbox, 38→4)
09:41 you’re back · 1 thing needs you
How it actually holds up
The freedom is the promise.
The engineering is the proof.
Anyone can ship a demo. The reason you can actually leave the desk is the unglamorous layer underneath — queues, retries, observability, budgets, and an on-call posture for the system, not for you.
Reliability
Idempotent steps, exponential backoff, dead-letter queues. Failures are events, not outages.
Observability
OpenTelemetry traces per run. Every tool call, token, and dollar is queryable.
Orchestration
Durable workflows (Temporal / LangGraph) so a 2 a.m. crash resumes from the last good step.
Guardrails
Per-agent budgets, rate limits, input/output validation. The agent can't run away with the credit card.
Deployment
Containerized, versioned, rollback-ready. Postgres for state, Redis for queues, K8s when it earns it.
On-call
Alerts that mean something. The system pages itself first — you only hear about what actually needs you.
stacktypescriptpythonpostgrestemporallanggraphrediskubernetesoteltwiliolivekitopenaimcp
Who I work with
Two kinds of clients.
One common thread.
Different starting points, same finish line: a system you can actually leave running.
Both have outgrown the demo. Both need it to actually work.
→ Operators
Founders, small firms, and busy teams who want their work automated and just need it to run.
→ Developers
Engineering teams architecting and deploying agents that operate autonomously and reliably.
Voice agent waveform — live call audio
What I build
Systems, not scripts.
Agent orchestration flow diagram
02
Autonomous workflows
Multi-step processes that run unattended on a schedule or trigger — pulling data, making decisions, and taking action across the tools you already use.
temporal · postgres · n8n · mcp
03
Agent infrastructure
The reliability layer for teams running their own agents: orchestration, retries, observability, and deployment that survives contact with production.
kubernetes · otel · langgraph · redis
Let's talk
Have an agent that needs to grow up?
Free 30-minute call. We'll figure out if it's a fit.