Fan Dian, founder of XSmart Technology, is an outlier in today’s smart hardware sector.

Fan was among Xiaomi’s founding employees and previously served as general manager of its internet-of-things platform. He also chaired Xiaomi’s artificial intelligence-of-things strategy committee during his tenure.

Over the past two years, hardware startups have largely followed a familiar playbook: raise capital on the strength of a founder’s resume, quickly build a product for crowdfunding, generate media attention, and use that momentum to raise more funding.

Fan took a different route. He spent three years building his first product.

His choice was also unusually niche. Instead of developing AI-driven wearables or mattresses, categories that already have market validation, he explored a different product type: a bedside lamp.

Little has been disclosed about what happened during those three years. Fan rarely accepts interviews or meets investors. That low profile extends to XSmart’s fundraising. In three years, the company has disclosed only one angel round, backed by Xiaomi, FutureX Capital, and other investors.

On May 19, XSmart launched its first lamp on Kickstarter under the Sleepal brand, priced at USD 449. The product drew questions as quickly as it drew attention. Is a bedside lamp worth that price? Did three years of work produce only a lamp? And if users already have smartwatches, who needs it?

Measuring sleep with millimeter-wave radar

Fan told 36Kr in an interview after the lamp’s launch that the startup idea came from his own health problems. He has obstructive sleep apnea, a condition in which the airway can become blocked during sleep, reducing the body’s oxygen intake. Even after eight hours of sleep, he would wake up with severe headaches.

Such problems are not rare, according to Fan. Sleep apnea affects about 15% of adults, he said, with prevalence rising to about 30% among people over 40.

Fan tried various sleep improvement methods, including CPAP devices, AI mattresses, and sleep belts. CPAP, short for continuous positive airway pressure, is a treatment that keeps the airway open during sleep. But Fan said many long-term options involve friction for users.

AI mattresses are expensive and difficult to install, Fan said. His team also considered AI ceiling lights, which can be designed to sense an entire room but may be obstructed by ceiling structures. Wearable sleep devices, including wristbands and rings, have been validated by the market, but Fan’s research found that only 60% of wearable owners wear them at night, while adoption among older users remains low.

Fan attributed that to several factors, including the need to keep wearables charged and discomfort from wearing them during sleep. XSmart therefore designed its product as a bedside lamp, aiming to reduce user friction.

That approach improves the user experience, but it also raises the technical bar.

Wearables typically measure sleep using optical PPG, or photoplethysmography, signals to detect capillary heart rate. They then combine those readings with wrist acceleration to estimate body movement and infer sleep status. This method can be affected by arm hair, tightness, skin tone, tattoos, and other factors. Fan said wearables also cannot fully capture a user’s sleep environment, including whether temperature, noise, or light is affecting sleep, or how a user’s sleeping posture changes.

Sleepal uses a different approach. Fan said transitions between sleep stages are governed by the central nervous system and autonomic nervous function. Physiological signals, such as breathing and heart rate, change together and are closely linked to sleep stages. Breathing signals are also a core basis for identifying apnea.

To collect this information, Sleepal uses a sensor matrix that includes a 60-gigahertz millimeter-wave radar, a thermal array sensor, a microphone array, and an environmental sensor. The millimeter-wave radar detects body movement during sleep and extracts breathing rate, heartbeat characteristics, and chest movement with accuracy down to 0.1 millimeters to infer sleep stages. The microphone array captures snoring and ambient noise, the environmental sensor detects indoor light and sleep disruptions, and the thermal array senses body contours to determine sleeping posture.

Fan said Sleepal uses contactless monitoring while continuously tracking direct vital signs, including breathing, heartbeat, and full-body movement. He said the resulting estimates are less error-prone than wearable methods that rely mainly on movement.

A paper co-authored by XSmart and Thomas Penzel, president of the World Sleep Society, validated Sleepal using 1,022 nights of hospital PSG data. PSG, or polysomnography, is a clinical sleep study used to monitor sleep stages and related physiological signals. Sleepal reportedly achieved a kappa score of 0.695, a measure of agreement between model output and reference labels, higher than the figures cited for the Apple Watch and Oura Ring.

Building models from hospital sleep data

Once the company addressed data collection, its next challenge was building a benchmark.

In sleep medicine, PSG is considered the gold standard. After data is collected, professional technicians manually label sleep stages such as wakefulness, light sleep, deep sleep, and REM (rapid eye movement), as well as apnea events. In model training, those labels serve as standard answers for supervising and aligning raw radar signals.

Fan said XSmart worked with multiple hospital sleep centers over three years to collect more than 2,000 nights of PSG data. Annual spending on data reached a seven-figure RMB sum, he added.

Using this benchmark data and user sleep data, the company trained seven specialized AI models: a vital sign detection algorithm, a multimodal sleep staging model, an apnea detection model, a multimodal human-state recognition model, a multimodal sleep posture recognition model, an on-device snoring recognition model, and a radar signal ECG generation model. ECG refers to electrocardiogram data, which records the heart’s electrical activity.

With this matrix of vertical models, Sleepal becomes more than a sleep data collection device. It can provide personalized improvement suggestions, such as identifying whether sleeping on one’s back worsened snoring, or whether micro-awakenings were caused by environmental noise, light, or heavy exercise before bed.

The company also added features built around sleep routines, including a circadian rhythm lamp, white noise, and a smart alarm. Before a scheduled alarm time, Sleepal gradually brightens and plays sound only when the user is in light sleep or a micro-awakening phase, reducing the abruptness of waking from deep sleep. When a user gets up at night, the lamp turns on dimly and switches off again when the user returns to bed.

In the first 48 hours after launching on Kickstarter, Sleepal raised more than USD 200,000, outperforming 90% of products on the platform, according to Fan.

Fan said Sleepal’s short-term business model will combine hardware sales with software subscriptions. In the long term, the company plans to expand from sleep detection into home-based health AI solutions. The multimodal sensing technologies validated in the bedroom could, for example, be applied in bathrooms for fall detection using millimeter-wave radar, or in dining areas for diet detection.

Fan ultimately wants continuous vital sign data to support chronic disease risk screening and prevention, turning the product into an entry point for at-home health management.

The bedside lamp is only its first step.

KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Qiu Xiaofen for 36Kr.