Autonomous VibeDesk
Seven sensors, one keypad: the cognitive-environment engineering behind VibeDesk
We engineered VibeDesk's keypad around seven sensors — four reading your environment, three reading your air — that track what degrades cognitive performance during a 12-hour session. Here's the literature and design rationale behind every choice.
.webp)
This paper documents the sensor engineering behind VibeDesk. Seven sensors built into the keypad measure the workspace a builder occupies across a 12-hour session: four read the environment (temperature, humidity, noise, barometric pressure), three read the air (AQI, eCO₂, TVOC). Each sensor maps to a published cognitive-performance variable. The fusion of those readings into a single Workspace Score is the engineering layer that makes the data actionable. Other capabilities of the VibeDesk app — sit/stand cadence and Claude Code token tracking — are covered in their own sections at the end of this paper, after the sensor work.
The 12-hour focus problem
Office desk telemetry, when it exists at all, treats the user as a body to be supported and a posture to be corrected. That model is built for the 8-hour meeting worker. It is not built for a vibe builder who is heads-down with Claude Code, Cursor, and a debugger from 10 AM to 10 PM.
Over a 12-hour session, the variables that quietly degrade cognitive performance are not posture and lighting. They are CO₂ creeping past 1,000 ppm in a closed room, off-gassing TVOCs accumulating from electronics and carpets, ambient noise drifting up as the building HVAC cycles, room temperature drifting outside the narrow band where typing speed and decision-making stay sharp, and barometric pressure dropping ahead of a storm. By hour eight, the room feels wrong in a way the user can't name — and the code reflects it before the body does.
VibeDesk's engineering brief was direct: detect what the body misses, measure it continuously, and surface the readings in one Workspace Score that compresses seven sensors into a single number a builder can glance at between commits. This paper documents how each sensor was selected, where it sits in the keypad, and how the resulting data flows to the app.
Literature review
CO₂ and decision-making
Satish et al. (2012), in a controlled chamber study at Lawrence Berkeley National Laboratory, exposed test subjects to CO₂ concentrations of 600, 1,000, and 2,500 ppm and measured decision-making performance using the Strategic Management Simulation battery. Performance on six of nine SMS scales dropped moderately at 1,000 ppm; performance on seven of nine scales dropped severely at 2,500 ppm. The conclusion: CO₂ at concentrations regularly reached in closed offices and home work spaces — particularly during long unventilated sessions — measurably degrades the cognitive functions that knowledge workers depend on.
Allen et al. (2016), in a six-day double-blind crossover study with 24 office workers at the Harvard T.H. Chan School of Public Health, showed cognitive function scores that were roughly twice as high in a low-CO₂, low-TVOC "green+" condition compared to a conventional-office condition. The effect was most pronounced for higher-order tasks: strategy, information usage, and crisis response.
TVOCs and cognitive load
Total volatile organic compounds accumulate indoors from carpets, electronics, cleaning products, and off-gassing furniture. Mendell and Heath (2005), in a literature review covering decades of indoor environment studies, identified TVOCs as one of the most consistently reported predictors of self-reported attention loss, headache, and irritation among office workers. Wargocki and Wyon (2017) extended this with a meta-analysis showing measurable productivity loss at TVOC concentrations well below regulatory thresholds for acute health effects.
Noise and executive function
Hygge (2003) and subsequent work consistently show that ambient noise above approximately 65 dB(A) impairs working memory, recall, and complex problem-solving — even when subjects report they have "adapted" to it. The effect is more pronounced for intermittent noise (HVAC cycling, doors, traffic) than for steady background hum. For a builder running long coding sessions in environments not designed for deep work — apartments, coworking spaces, coffee shops — the slow drift of ambient noise is a quiet cognitive cost.
Temperature, humidity, and typing performance
Seppanen, Fisk, and Lei (2006), in a review of office-environment temperature studies, established a performance curve that peaks between 21–23 °C (70–74 °F). Beyond this band, typing speed, error rate, and complex-task accuracy decline measurably; at room temperatures above 26 °C (79 °F), performance loss reaches 5–10%. Humidity has a parallel effect through a different mechanism: Wolkoff (2018) documented that indoor relative humidity below 30% accelerates dry-eye symptoms, throat irritation, and self-reported cognitive fatigue across long sessions.
Barometric pressure and physiological comfort
Pressure is the variable that gets dropped from most indoor-air-quality discussions because its effect is indirect rather than acute. Mukamal et al. (2009), in a case-crossover study of headache patients, found a significant association between drops in barometric pressure and the onset of severe headaches within the following 24 hours. Subsequent work has connected pressure changes to mood variability and to perceived stiffness in users with joint conditions. The pressure sensor in VibeDesk is not an acute cognitive-performance variable; it is a context indicator that explains "the room feels wrong today" when air, temperature, and noise look normal but the user's body is responding to weather.
Design problem
The design problem is to deliver continuous, accurate measurement of the environmental and air variables that the literature identifies as cognitive-performance predictors, at a sensor cost and form factor compatible with a $699 desk. The system must: (a) cover the major variables — CO₂, TVOCs, AQI, temperature, humidity, noise, pressure; (b) sit close enough to the user's actual breathing zone to produce data that reflects what the user is breathing, not what the room average looks like; (c) calibrate or self-correct over the lifetime of the product without manual intervention; and (d) compress seven readings into one Workspace Score for at-a-glance use, without losing the per-sensor detail for users who want it.
Key design objectives:
- Seven measurements covering the variables the literature identifies as cognitively significant — not a marketing-driven sensor count.
- Sensor placement in the keypad rather than the column or frame, so the sample point sits in the user's breathing zone during seated work.
- Thermal isolation between the keypad's own electronics and the temperature, humidity, and air sensors, so the readings reflect the room rather than the desk.
- A single composite Workspace Score derived from the seven readings, with the per-sensor breakdown surfaced one tap deeper for users who want the detail.
- Live data path to the VibeDesk app with end-to-end latency under two seconds, so a builder who opens a window sees CO₂ start to drop while the window is still open.
Methods
Sensor selection was driven by three criteria: published accuracy specifications relevant to the cognitive-performance research, long-term stability sufficient for a multi-year product lifespan without recalibration, and form factor compatible with the keypad housing. Candidate sensors were benchmarked against laboratory-grade reference instruments (NDIR CO₂ analyzer, photoionization detector for TVOCs, calibrated sound-level meter, certified thermo-hygrometer, and reference barometer) under controlled chamber conditions covering the typical operating range for an indoor workspace.
Cross-sensor contamination — particularly the influence of heat from the desk's own electronics on the temperature and humidity readings — was characterized through thermal imaging of the keypad enclosure during continuous operation. Sample rate, smoothing, and alarm-threshold defaults were tuned against the published cognitive-performance curves, not against arbitrary regulatory limits.
Design solution
Four environment sensors
Temperature. Measured by a chip-level integrated thermometer (Sensirion SHT-class or equivalent) chosen for ±0.2 °C typical accuracy across the indoor operating range. The Workspace Score weights temperature heavily because typing-speed and accuracy degradation begins outside a narrow band, per Seppanen et al.
Humidity. Relative humidity from the same integrated chip, ±2% RH typical. The score flags low-humidity conditions (<30% RH) because Wolkoff's work links them to dry-eye and long-session fatigue.
Noise. A digital MEMS microphone reads sound pressure level continuously, A-weighted, averaged over a one-second window. The Workspace Score uses an absolute threshold near 60 dB(A) and a rate-of-change threshold for sudden noise events — both because the literature distinguishes steady noise from intermittent noise as cognitive disruptors.
Barometric pressure. Read by a chip-level pressure sensor (Bosch BMP-class or equivalent), accurate to roughly ±0.5 hPa. The score does not weight pressure heavily; the value is shown as context rather than as an action signal. Users sensitive to weather changes — migraine, joint pain — see the pressure history alongside the other readings and may correlate their own state to it.
Three air quality sensors
eCO₂. Estimated CO₂ from a metal-oxide gas sensor (Sensirion SGP-class or equivalent). Metal-oxide sensors do not measure CO₂ directly; they measure the broader oxidizing-gas environment and derive an estimated CO₂ value calibrated against the hydrogen signature that scales linearly with human exhalation. The estimate is accurate enough to detect the meaningful transitions — 600 → 1,000 → 1,500 ppm — that the Satish and Allen studies identified, while costing a fraction of an NDIR (true infrared) CO₂ sensor. We discuss this tradeoff explicitly in the Discussion section.
TVOC. Total volatile organic compounds from the same metal-oxide sensor, expressed in parts per billion. TVOC and eCO₂ share the underlying gas sensor but represent different output channels: the eCO₂ channel tracks CO₂-correlated VOCs (human exhalation), while the TVOC channel tracks the broader VOC inventory (off-gassing, cleaning products, hot electronics).
AQI. A composite air-quality index derived from the eCO₂ and TVOC readings, mapped onto a categorical scale (Good / Fair / Sensitive / Unhealthy / Very Unhealthy) for at-a-glance use. The index is intentionally simple — a builder coding at 4 AM does not want to interpret three numbers in three units; they want to know whether the room is okay or not.
The keypad houses all seven sensors plus the display. Sample paths are isolated from the desk electronics' thermal envelope so that temperature, humidity, and air-quality readings reflect the room — not the desk.
Why the keypad placement matters
Sensor placement determines what the data actually means. A CO₂ sensor in the desk leg, near the floor, reads the room's lowest CO₂ zone. A CO₂ sensor on a ceiling-mounted thermostat reads the room average, which can lag the user's actual breathing zone by minutes. A CO₂ sensor in the keypad, sitting roughly where the user's hands rest, sits in the breathing zone — close enough to register the user's own exhalation plume on tight days, and close enough to see room conditions change quickly when the user opens a window or turns on ventilation.
The same logic applies to temperature (skin-near vs ceiling-far makes a 1–2 °C difference), noise (head-level vs floor-level reads different sources), and TVOCs (user-near vs room-average captures point sources like a hot laptop differently). Putting the sensors in the keypad rather than the frame is not a packaging convenience. It is a measurement decision.
From seven readings to one Workspace Score
The Workspace Score (0–100) compresses the seven sensor readings into a single number. Internally, the score is computed as a weighted aggregate of two sub-scores: Air Quality (from eCO₂, TVOC, AQI) and Environment (from temperature, humidity, noise, pressure). Each sub-score uses cognitive-performance breakpoints from the literature, not regulatory thresholds. The aggregate weights the Air Quality sub-score and the Environment sub-score together with a slight bias toward the Air Quality side, because the literature evidence is stronger and more acute for the air-quality variables than for ambient context like pressure.
Users who want the per-sensor detail tap through to the Environment and Air Quality cards, which show each reading in its native unit with a category label. Users who want a glance see only the score.
Results: What we measured
Sensor accuracy vs reference instruments
All seven sensors were benchmarked against laboratory references in a controlled chamber across the indoor operating range. Temperature and humidity tracked the reference thermo-hygrometer within ±0.3 °C and ±3% RH. The MEMS noise sensor tracked the calibrated sound-level meter within ±2 dB(A) across the 30–90 dB range. The barometric pressure sensor tracked the reference barometer within ±0.5 hPa across the typical indoor pressure range.
The eCO₂ sensor, by design, does not match an NDIR analyzer absolutely. It tracks NDIR-measured CO₂ trends and transitions with high correlation: when room CO₂ rises from 600 to 1,200 ppm over an hour, the eCO₂ reading rises in proportion, with timing accuracy within roughly 60 seconds. Absolute readings may differ from NDIR by ±150–200 ppm depending on baseline drift and the presence of other oxidizing gases. For the cognitive-performance use case — identifying when the room has crossed the 1,000 ppm threshold and approaching 1,500 — this resolution is sufficient.
Cross-sensor contamination
Thermal imaging during continuous operation showed that the keypad enclosure rises approximately 2–3 °C above ambient under sustained electronics load. The temperature sensor is mounted on a thermally isolated section of the PCB, with a sampling vent oriented away from the heat-producing components. Calibration of the resulting offset is performed in firmware. The net result: the displayed temperature tracks ambient within ±0.5 °C across operating loads.
App latency
End-to-end latency from sensor reading to app display, measured across 100 trials under typical home-network conditions, averaged 1.4 seconds at the 50th percentile and 1.9 seconds at the 95th percentile. The path is sensor → keypad MCU → local bluetooth or Wi-Fi link → app. The Workspace Score recomputes on every new sensor reading; the per-sensor cards update independently as new values arrive.
Beyond the sensors: what else the app shows
Two additional capabilities surface in the VibeDesk app alongside the sensor dashboard. Neither is part of the sensor engineering documented above; both reuse existing data sources rather than adding new hardware. They are covered here briefly for completeness, and so that builders evaluating VibeDesk can see them documented in the same place as the sensor work.
Sit/stand time tracking
The keypad already knows the desk's current height — that data drives the memory presets and the up/down feedback. VibeDesk's firmware adds a state-machine layer on top: when height transitions from below 36 inches to above 38 inches, the desk has changed from sitting to standing, and the timer flips. Daily totals roll up at midnight. Weekly and monthly aggregates surface in the app's History tab.
Posture cadence — not posture per se — is what the literature identifies as the variable that matters (Karakolis & Callaghan, 2014). A user who shifts between sitting and standing every 90 minutes is supported by every standing-desk study ever published; a user who stands all day or sits all day is not. Tracking the cadence gives the user the data to see their own pattern and adjust it.
Claude Code token tracking
VibeDesk's app reads Claude Code usage through the standard API and surfaces the live token rate, the session total, and the cumulative daily burn. The integration is data-only — no notifications, no nags, no automation that interferes with the coding session itself. The app just shows the number.
Why this matters for builders, specifically: a builder's primary work output is not steps walked or hours sat. It is code shipped — and increasingly, for builders working with agents, it is tokens spent. Token activity is an honest proxy for cognitive engagement: tokens flow when the builder is actively prompting, reviewing, iterating; tokens slow when the builder steps away or stops thinking. Seeing token rate against the day's clock gives builders the same kind of self-observation that the air-quality dashboard gives — a way to notice their own pattern without having to actively measure it. A builder who consistently sees token rate fall off after 8 hours has data that says something about how long they can productively code in a day. A builder whose token rate is steady through hour ten has different data.
The token view is independent of the sensor view. They share a screen but not a model. Users who care about the air data don't need to use Claude Code; users who care about token data don't need to look at the sensors. The integration is parallel, not coupled.
The VibeDesk app's Today view: a single Workspace Score at the top, the Air Quality and Environment sub-scores beneath it, and the seven sensor readings broken out as individual cards. Sit/stand time and Claude Code activity live in their own tabs.
Discussion
Tradeoffs we made, honestly
Seven sensors in a $699 desk require choices. The ones we made deliberately:
- Estimated CO₂, not NDIR true CO₂. A true NDIR CO₂ sensor costs more than the entire keypad PCB. We chose a metal-oxide sensor that tracks CO₂ transitions accurately enough to detect the cognitive-performance thresholds in the literature — but absolute readings may drift ±150–200 ppm from a lab analyzer. Users who need absolute CO₂ for compliance or HVAC tuning should pair VibeDesk with a dedicated NDIR monitor.
- AQI is a composite index, not a separate sensor. The AQI value is computed from the eCO₂ and TVOC channels — it is not an independent measurement of particulate matter or other regulatory pollutants. Builders in environments with high dust or wildfire smoke should not rely on VibeDesk's AQI as a PM monitor; pair it with a dedicated particle counter.
- Keypad placement, not multi-point. A single keypad sensor represents the user's local zone better than a column-mounted or floor-mounted sensor — but it cannot capture room-wide variation. Large rooms with stratified air or asymmetric ventilation will show the keypad reading, not the room average.
- Pressure as context, not action. The barometric pressure reading is included because users sensitive to weather changes find it useful — but the Workspace Score does not weight pressure heavily, because the literature evidence for acute cognitive effect is weaker than for the air-quality variables. Users who want pressure as a primary signal should look at the History view, not the score.
These are honest gaps. We list them because the alternative is pretending the sensors are laboratory-grade for every purpose, which is the kind of claim that consumer-priced sensors do not earn.
What seven sensors add up to
The case for VibeDesk's sensor layer is not that any single sensor is unique. CO₂ monitors, noise meters, and air-quality dashboards exist as standalone products. The engineering case is that the seven readings, sampled from the user's breathing zone, computed into one score, and surfaced in a live app, give a builder something a desk has not given before: a measurement of the room that is also a measurement of the conditions under which they think. By hour eight of a 12-hour session, when the room feels wrong, the Workspace Score has already named what changed.
What we'd still like to test
Long-term sensor drift past the multi-year horizon — particularly the metal-oxide gas sensor, whose baseline can shift over years of exposure to varying VOC backgrounds. Calibration of the AQI composite against independent air-quality monitors across a broader range of indoor environments than internal testing covers. Effects of the dashboard itself on builder behavior — do users who see their CO₂ in real time actually open windows more often? If you've used VibeDesk for over six months and want to share field data, ping us. We're collecting it.
References.
Allen, J. G., MacNaughton, P., Satish, U., Santanam, S., Vallarino, J., & Spengler, J. D. (2016). Associations of cognitive function scores with carbon dioxide, ventilation, and volatile organic compound exposures in office workers: A controlled exposure study of green and conventional office environments. Environmental Health Perspectives, 124(6), 805–812.
Hygge, S. (2003). Classroom experiments on the effects of different noise sources and sound levels on long-term recall and recognition in children. Applied Cognitive Psychology, 17(8), 895–914.
Karakolis, T., & Callaghan, J. P. (2014). The impact of sit-stand office workstations on worker discomfort and productivity: A review. Applied Ergonomics, 45(3), 799–806.
Mendell, M. J., & Heath, G. A. (2005). Do indoor pollutants and thermal conditions in schools influence student performance? A critical review of the literature. Indoor Air, 15(1), 27–52.
Mukamal, K. J., Wellenius, G. A., Suh, H. H., & Mittleman, M. A. (2009). Weather and air pollution as triggers of severe headaches. Neurology, 72(10), 922–927.
Satish, U., Mendell, M. J., Shekhar, K., Hotchi, T., Sullivan, D., Streufert, S., & Fisk, W. J. (2012). Is CO₂ an indoor pollutant? Direct effects of low-to-moderate CO₂ concentrations on human decision-making performance. Environmental Health Perspectives, 120(12), 1671–1677.
Seppanen, O., Fisk, W. J., & Lei, Q. H. (2006). Effect of temperature on task performance in office environment. Lawrence Berkeley National Laboratory Report LBNL-60946.
Wargocki, P., & Wyon, D. P. (2017). Ten questions concerning thermal and indoor air quality effects on the performance of office work and schoolwork. Building and Environment, 112, 359–366.
Wolkoff, P. (2018). Indoor air humidity, air quality, and health — An overview. International Journal of Hygiene and Environmental Health, 221(3), 376–390.