Building Atmospheric Intelligence

Six sensory channels, a benchmark engine, a deterministic AI inference pipeline, and a data moat that compounds with every audit. Here's what the build looks like from the inside.

There is a specific conversation I kept having with hotel GMs and restaurant operators in Miami. A space is underperforming. The metrics point to it: dwell time falling, NPS scores softer than the offer warrants, a bar that fills early and empties before midnight. The team is performing. The concept is right. Something in the space itself is producing a result that nobody can trace back to a cause.

The response, in most of these conversations, was to adjust what was visible - the concept, the programming, the design language. Understandable. These are the variables that appear in a brand review and can be actioned in a budget cycle. What they could not address is the layer underneath: the sensory environment that acts on a guest before they have consciously registered anything. Light quality, acoustic behaviour, scent, air, the thermal character of the room. These conditions shape how people feel, how long they stay, and whether they return - and they have been largely invisible to the instruments operators actually use.

That connection, between what a space does atmospherically and what it produces commercially, - is the territory Sensoria is built to make visible. Building the platform to do that required a set of architectural decisions that, from the outside, can look like constraints. They are not constraints, however. They are the correct choices for what the product needs to do.

Sensoria is a self-serve SaaS audit platform. An operator runs a structured assessment of their space across six sensory channels: light, acoustics, scent, thermal and air quality, tactile surfaces, visual field, plus a behavioral layer and a business context calibration. Each channel is scored against a benchmark engine fine-tuned to the space type and concept positioning. The output is an intelligence report: what the space is doing across each channel, where the gaps are, what the specific interventions are, and what changing them is likely to produce commercially.

The most important early architectural decision was how the AI layer works.

The platform uses the Anthropic API as a structured inference engine rather than a language model - not to reason freely or generate suggestions, but to score structured inputs against calibrated benchmarks and return validated outputs. The same inputs always produce the same outputs. Every recommendation is traceable to a benchmark model, not to a generative process.

Here is why that matters in practice. When the platform identifies that a hotel bar is running its lighting at 4000 Kelvin and recommends shifting to 2800 Kelvin, that recommendation traces back to a specific mechanism: what that colour temperature shift does to the retinal dopamine pathway at evening occupancy, against the benchmark data for that space type. The recommendation is the output of a model encoding a precise scientific relationship. It is consistent, auditable, and traceable to its source.

This is what it means to use AI as a calibration instrument rather than a generative engine, and it required building the inference pipeline with that precision from the start. Every AI agent in the system is programmed to operate within a tightly defined scope: receive structured input, score it against the relevant benchmark model, return a validated output. No free reasoning. No probabilistic variance. The calibration of the system is what makes the recommendations actionable rather than approximate.

Building a platform that reasons this precisely about physical spaces also required rethinking how the audit itself is structured. A linear audit - the same sequence of questions for every space - produces consistent data but misses the diagnostic intelligence that comes from context. If a space has significant acoustic problems, the questions that follow should respond to that. A branching audit adapts: each answer shapes what comes next, the way a genuinely diagnostic conversation works. It is significantly more complex to build, and it is the correct architecture for the product I am building. Constructing the branching logic at v1, rather than retrofitting it later when the behavioral input layer arrives at v1.5, was one of those decisions that only matters in year two - and that determines whether year two is an expansion or a rebuild.

The benchmark engine is where the genuinely hard work lives.

The scientific foundation spans environmental neuroscience, psychology, hormonal biology, and sensory biology - four disciplines that together map the chain from physical stimulus to human response. Light wavelength affects retinal cell activation and downstream hormonal state. Acoustic energy levels correlate with cortisol and cognitive load. Scent compounds reach the limbic system through a pathway that bypasses conscious processing entirely. These are not design opinions. They are documented mechanisms, and the benchmark models encode them with enough precision to produce specific, actionable recommendations rather than general guidance.

Those benchmarks are adjusted to space type, concept positioning, and target demographic. A luxury hotel lobby operates under different standards from a fast-casual restaurant or a wellness studio: different lighting ratios, different acoustic tolerances, different scent intensity thresholds. Building these models from the scientific data and then validating them against real-world audit data is work that cannot be shortcutted. And it does not reach a fixed endpoint: every audit adds a data point, every space type becomes more precisely understood, every new context makes the benchmark engine more specific to the spaces and the operators it is built to serve.

In the same way every audit refines the benchmarks. The accumulated dataset, (atmospheric conditions correlated with behavioral and commercial outcomes, across space types and geographies), becomes more precise with every client. It cannot be replicated without running the same audits. It deepens from the moment the first audit runs.

The product is the entry point. The benchmark database compounding behind it is the intelligence layer I am building - one that gets smarter with every client, and a data asset that becomes increasingly valuable and increasingly irreplicable as the audit volume grows.

v1 ships with the full six-channel audit, space onboarding, benchmark engine, and intelligence report. The goal for this phase is proof: a small number of operators who learn something they did not know, act on the recommendations, and see the change in their business metrics. The atmospheric lever is there in every space. The instrument to read it is being built now.