WHAT
Every tool in the stack could tell us how a building performs. None of them would say how a room feels. Sensi closes that gap: a copilot that reads a floor plan and scores comfort across six senses (thermal, visual, acoustic, spatial, olfactory, tactile), calibrated to one person at a time, not an average. It was prototyped first as an MCP tool, then rebuilt as a standalone app. Project lead, A to Z, built by a team of four: Lakzhmy Mari Zaro, María Sánchez Domínguez, Charles Abi Chahine and me.
WHY
Comfort is usually the thing that shows up, or does not, after the design is done. Sensi makes it a layer you can interrogate while the plan is still soft, because you do not walk into a room and average your experience: the thing that is wrong is the thing you notice. And nothing you fix stays fixed alone: the whole project is the ripple, what a change drags along.
HOW
- Onboarding calibrates it to one person: thermal grudges, noise tolerance.
- One LLM call per turn routes each request through a LangGraph graph: analyze, edit, preview, audit. Two evals ride inside: a critic on every suggestion, an evaluator on every reply.
- A coupling matrix ripples every change into the neighboring senses, so a fix that quietly breaks another score gets flagged, not hidden.
- Edits preview before they commit, then a vision model redraws the room, structure intact, and hands over the report.
WHAT CAME OF IT
Two LLM providers scored the same three scenes, one arranged to fail, and mostly agreed. Easy to call that validation. We wrote agreement is not truth into the notes instead, and kept every disagreement as data. It hardened into evals: swap a model and the same apartment replays through the same prompts, judged blind before the swap is believed. Nobody has stood in a room Sensi scored and said whether it was right. It estimates, it does not measure, and every coupling in the code is labeled verified or inferred, so you can see which links rest on research and which on our reasoning.