Why it is built this way.

India's outdoor workforce — construction, agriculture, delivery, street vending — cannot step into an air-conditioned room between shifts. The ILO projects that heat stress will cost India about 5.8 % of working hours by 2030, roughly 34 million full-time jobs (ILO 2019, as reported by Down To Earth). The 2025 Lancet Countdown estimates 247 billion potential labour hours lost in India in 2024 (Lancet Countdown 2025).

Vellore is a hard case. May's mean daily maximum is about 39 °C and the record is 45 °C, with afternoon humidity around 43–51 % in the hot months (IMD normals relayed by Wikipedia). That sits right at the edge of where a fan stops helping.

Policy is moving the same way. India's NDMA draft Guidelines for Cooling Centers (May 2025) call for temporary shaded structures for outdoor workers at work sites, with high-reflectance roofs, no bare GI sheet roofing, shaded or green unpaved ground and cross-ventilation (NDMA 2025). SAAYA is a testable, low-cost version of exactly that structure, plus a way to prove whether it works.

A normal cooling product is judged by how cold it gets a closed room. An outdoor rest spot has no walls to hold cold air, so that yardstick does not apply. SAAYA attacks the larger part of outdoor heat load — radiation from the sun, sky, roof and hot ground — and uses a small fan only where it helps. Its genuinely new part is not the shelter; it is a control and evidence layer that knows when a fan is not spinning, when a sensor is lying, and when a comparison is not valid.

We wrote a small steady-state model (api/thermal.py, 17 test cases) instead of quoting a temperature drop. It balances sun, sky, wind and grey-body radiation on each roof skin, then computes the mean radiant temperature (MRT) at a seated worker's head from the exact solid angle of the roof, sky and ground. MRT is what the body feels from surrounding surfaces; it is the quantity a shelter actually changes.

Loading model…

What this changed in the design

The roof is only a small part of what a seated person sees; sky and sunlit ground fill most of the rest. So once the sun is blocked, roof finish moves the seated radiant load by a couple of degrees, while overhang depth and what the ground around the shelter is made of move it by several. That reordered our priorities: deep overhang and ground treatment first, a cool roof lining second, and the second roof skin as an optional, removable experiment rather than a core cost.

One more surprise: we had specified a bare, low-emissivity metal lining. Once the model counted the hot sunlit ground the lining reflects, a painted (high-emissivity) lining came out marginally better. What matters is that the lining stays cool, not its finish — so the cheaper, more corrosion-resistant painted lining is acceptable.

Each design is reported at its less favourable seat: in mid-May the noon sun is north of the zenith at Vellore's latitude, so the shadow shifts south and the rear bench is the harder case.

A fan cools by evaporating sweat. When the air is hotter than skin and very dry, it can also blow heat onto the body faster than evaporation removes it. The evidence:

• In 8 young men at 42 °C, a fan still helped, but only up to a critical humidity — heart rate started rising at 47 % RH with a fan versus 38 % without (Ravanelli et al., JAMA 2015).
• In 9 adults aged about 68 at 42 °C, the fan gave no benefit; heart rate and core temperature were slightly higher with it (Gagnon et al., JAMA 2016).
• In a very hot, dry condition the fan made every measured outcome worse; in a hot, humid one it helped (Morris et al., Annals of Internal Medicine 2019).
• A biophysical model proposes fan-use ceilings of 39 °C for young healthy adults, 38 °C over 65, and 37 °C for older adults on anticholinergic medicines, and advises against routine fan-only advice in northern India (Morris et al., Lancet Planetary Health 2021).
• WHO's current fact sheet says to use fans only below 40 °C (WHO).

What SAAYA does with that

The firmware now screens fan benefit by air temperature: below 37 °C the fan is expected to help; from 37 °C the display says the benefit is limited for older or medicated people; from 39 °C automatic mode stops and the display reads “Hot: fan may add heat”, restarting only below 38.5 °C. Manual mode is never blocked — the worker decides, the shelter informs. The rule is identical in C++, Python and the browser, and tested in all three. It is a conservative screen built on the strictest published ceilings, not a clinical rule; humidity-dependent limits need the full Morris 2021 tables, which we have not yet extracted.

This is the clearest difference from a normal fan controller, which is a thermostat: it switches on harder as it gets hotter, which is exactly backwards above about 39 °C.

A cooling product usually reports a number. SAAYA reports a number only when it can defend it, and says why when it cannot. The rules are fixed before any data arrive:

• Predeclaration. The endpoint, which direction counts as better, and the smallest meaningful difference are locked into the record before the first sample. Changing them afterwards is impossible through the API. This is the same idea as preregistering a clinical trial: it stops us choosing the analysis after seeing the result.
• A simultaneous reference. Every intervention reading is paired with a matched reference reading taken at the same second. Weather changes affect both.
• Invalidation, not silence. Mixed sources, stale data, a fan that was commanded but did not spin, a mismatched reference or a gap in the record each produce an explicit “invalid” with the reason.
• One run is never a result. A single run reports only whether its point estimate lies beyond the margin, in either direction, or within it. The margin is declared in advance and set above what the instruments could produce on their own: two probes rated ±0.5 °C could disagree by up to about 1 °C before the co-location swap removes their fixed offset, so we declared 1.0 °C.
• No false precision. Sixty readings one second apart are not sixty experiments. The record reports the lag-1 autocorrelation and the effective sample size to show why no within-run confidence interval is given.
• Replication across days. Uncertainty comes only from independent runs, and runs on the same afternoon share weather, so the system requires at least three runs on three distinct days. It computes a two-sided 95 % Student-t interval of the run means (stricter than the 90 % interval of a standard equivalence test) and claims a direction only if the whole interval clears the margin, or “no meaningful difference” only if the whole interval sits inside it. Invalid runs are reported, never dropped.

Bench prototype: from ₹10,409 to about ₹4,661

Two premium parts make up most of the current bench cost: two Adafruit SHT31 carriers (₹3,066) and an imported Noctua fan (₹3,130). A costed budget profile in research/costs.json replaces them with GY-SHT30 modules (₹179 each) and a generic 4-wire PWM fan with a tachometer, and uses a ₹49 DS18B20 probe (Probots). That brings the bench to about ₹4,661 — close to the original ₹2,150–4,200 dossier estimate — without giving up anything the evidence chain needs.

What we refuse to cut: the tachometer (cheap Indian 40 mm fans are 2-wire with no feedback, which would blind the fault detection), the second matched sensor for the paired reference, and calibration time. Removing those would make the system cheaper and unable to tell the truth.

Full-size shelter: spend where the model says it pays

• Make the upper roof skin optional. The model gives it 0.3–0.8 K at the seat. It is roughly half of the ₹9,000–16,000 roof-materials allowance.
• Treat the ground before buying more roof. Irrigated ground cover lowers seated MRT by about 4 K versus dry ground at no structural cost; 0.3 m more overhang on every side buys about 1.4 K but grows the roof by about half. Light paving makes it about 5 K worse.
• Paint the lining. A painted lining performs as well as bare low-emissivity metal at the seat and resists corrosion better.
• Retrofit existing shade first. Foundations and structural steel are the largest lines. Where a sound structure already exists, re-roofing its underside and adding the controller is the cheapest option and is the comparison we plan to run.
• Bolt repeated sections. The frame uses repeated tube sizes and flat plates so cutting, drilling and jig costs fall with quantity.

Running cost

A 1.5-ton AC at about 1.4–1.6 kW uses roughly 11–13 kWh in an 8-hour shift. A 35 W BLDC fan uses about 0.3 kWh. The bench fan uses a few watt-hours. The shelter's real running costs are cleaning, repainting the reflective top and replacing fans and sensors; the annual maintenance allowance is ₹1,500–3,500.

• No shelter has been built and no temperature has been measured. The model predicts; it has not been validated.
• Mean radiant temperature is not heat stress. A worker's strain depends on humidity, work rate, clothing, hydration, acclimatisation and health. SAAYA never labels a person safe or recovered.
• The full-size cost is an allowance, not a quote. A dated Vellore fabrication quote and an installed AC quote are both still needed.
• Structural, electrical, foundation and weather design need qualified review before anyone sits under a full-size build.
• Novelty is limited and stated as such: shade, reflective roofs and fans are prior art. The contribution under test is the fault-aware evidence chain integrated with a low-cost shelter.