Michel Lopez MiscMich

Michel Lopez

Operator-builder and design engineer in Miami, working on production AI agents and hospitality systems.

LinkedIn · Email · Villa Paraiso Vacation Rentals

I like building software for work that has consequences. A wrong answer can become a customer problem, a missed task, or money moving without the right approval.

I was a Founding Product Engineer at Trellis (YC P26), where I worked across product, customer deployments, integrations, and reliability. I was also the design engineer for the product UI and landing page. I care about making a complex system feel clear, not merely making it function.

I also founded and run Villa Paraiso Vacation Rentals in the Florida Keys. That second job keeps the first one honest.

Some days I am debugging TypeScript. Other days I am figuring out why a cleaner cannot see a task or why two systems disagree about the same reservation. Most of my work lives somewhere in between.

What I build

Agent workflows with approvals, retries, recovery, and audit trails
Evals that replay failures instead of grading a polished demo
Production interfaces that make state, ownership, and failures legible
Landing pages and product surfaces that explain value without hand-waving
Local-first systems I can inspect when something goes sideways

What I care about

Clear ownership when automation touches a business
Deterministic code where an LLM does not need to make the call
Design that helps an operator understand what happened and what to do next
Customer context close to the people writing the software
Honest open source: useful, configurable, and clear about its limits

From Trellis to Villa Paraiso

My part at Trellis

As part of the founding product team, I contributed across:

Multi-channel agent interactions and durable execution
Turning production traces into replayable evaluations
Workflow controls for approvals, pausing, recovery, and inspection
Product UI, the public landing page, early visual direction, and launch messaging
Hospitality integrations and field debugging with customers

What I applied as an operator

Running Villa Paraiso gave those lessons a second test:

Operational work needs a visible state, a named owner, and a clear handoff
Consequential writes need a dry run, approval, or recovery path
Evidence should update a record; uncertainty should become a question
Empty, loading, and failure states should explain the operator's next move
Public tools should be useful while staying honest about what remains private

One recurring production failure shaped a lot of my work: an agent could produce a good answer while the underlying task was still unresolved. The lesson was that output is not completion. I carried that into Villa Paraiso's tooling through guarded writes in Guesty CLI, explicit state and recovery in the agent control plane, and trace-driven regression cases in production agent evals.

The same applies to design. A production interface should make status, ownership, evidence, and the next action obvious. That principle now shapes both the software and the visual systems I build.

Start here

Agent operations control plane

A dependency-free TypeScript control layer for long-running agents: approval gates, durable checkpoints, bounded retries, recovery, redaction, and append-only audit logs.

Release · Source

Production agent evals

A standard-library Python harness for replay cases, tool-call assertions, approval policies, structured output checks, and readable regression reports.

Release · Source

Guesty CLI

A working Python CLI for querying and synchronizing Guesty PMS data, with structured output, local caching, dry runs, and safeguards around writes.

OpenClaw Dashboard

A local-first dashboard for browsing an agent workspace, durable memory, filesystem activity, schedules, and optional gateway sessions.

Experiments and public reference builds

OpenClaw Mission Control explores workflow approvals, retries, leases, messaging, and audit trails.
RentalRadar is a demo-safe reference for monitoring brand visibility in AI-generated answers.
Loop is an experimental social playground with doodles, moods, voice notes, photos, small games, and an iOS widget.

These public editions use fictional or local data where the private product depends on customer systems.

Tools I reach for

Beyond the UI stack, I have used LangGraph for agent workflows, Trigger.dev for background work, and Vercel and Fly.io for delivery.

I am usually working on the boundary between an agent saying "done" and the business being able to prove it.

Navigation Menu