Embedded.fm · 2 Nov 2016 · 1h 09m
Evan Shapiro, CTO and co-founder of Knit Health, explains how his company uses a custom Linux-based camera with an onboard Qualcomm Snapdragon processor and deep learning neural networks to monitor infant sleep and respiration without wearable sensors. He details the engineering challenges of extracting weak respiration signals from noisy video, the decision to process data locally on the edge to preserve privacy and reduce cloud costs, and the rigorous security measures implemented to prevent hacking. Shapiro also discusses the history of neural networks, the 'Wizard of Oz' method used to train their models before having sufficient data, and his background as a professional poker player.
Where this conversation describes AI being used to build or deliver a product. Part of the AI in practice tracker.
Engineering & simulationdeep learning neural network
The AI model was trained to distinguish between a child and stuffed animals, locate body parts, and extract respiration signals from video frames to determine sleep states.
“It's got a deep learning neural network in it that's capable of understanding the difference between a child and a stuffed animal. It can recognize the child in the scene, know where they are, know where different parts of their bodies are.”
Speaker 3 · Hear it at 8:44
Engineering & simulationunsupervised learning mechanisms
The system uses unsupervised learning to tune itself to specific cribs, allowing it to generalize detection of objects like pacifiers or specific stuffed animals without manual retraining for every child.
“We're building in mechanisms where it can, uh, tune itself, essentially, where we can do some unsupervised learning so that if it's fairly confident, something's a certain stuffed animal or pacifier, in a given crib, um, it can change its, it can change the way it describes that concept to itself to make itself more confident…”
Speaker 3 · Hear it at 28:23
early-stage companies
Other shows on the same ground, closest first.
Anushka Salinas, CEO of Nanit, explains how the company evolved from a sleep-focused camera into a comprehensive smart parenting platform leveraging computer vision and AI. She details the technical architecture, including edge computing for critical breathing monitoring and cloud processing for long-term milestone tracking, while highlighting the shift toward a unified app ecosystem. The discussion covers the product's ability to reduce parental mental load through customizable smart alerts and the integration of audio monitors to create a cohesive family wellness solution.
David Wang, CTO of Elemind, explains how his company built a wearable headband that uses real-time EEG sensing and closed-loop audio stimulation to help users fall asleep faster. The device achieves a critical 2-millisecond latency loop by optimizing embedded firmware on a Cortex-M33 processor, avoiding expensive FPGAs while handling signal processing and edge AI for sleep staging. Wang details the 5-year development journey, including the shift from uncomfortable hair-penetrating electrodes to comfortable dry carbon-graphite sensors and the discovery that alpha wave suppression is more effective than boosting delta waves. The team is now developing an 'AI sleep tailor' to personalize stimulation for the 24% of users who do not initially respond to the standard protocol.
Perceive Founder and CEO Steve Teig explains how his company is bringing data-center-class machine learning to edge IoT devices through their new chip, Ergo. He details the decision to build a custom silicon and software stack from first principles to achieve 20 to 100 times greater power efficiency than existing solutions, enabling battery-powered devices like doorbell cameras to perform complex inference locally. Teig discusses the strategic choice to keep raw data on-device to address privacy and security concerns, contrasting Perceive's approach with cloud-dependent competitors that suffer from high latency and false positives. The episode highlights Perceive's early-stage commercialization, noting they are currently in production with several unnamed consumer and enterprise customers.
Chris Catterton, head of solution engineering at Micro AI (formerly OneTech), explains how their Adam ML technology enables lightweight AI training and inference on resource-constrained microcontrollers at the extreme edge of IoT. He details how this approach solves high data transmission costs by processing data locally and only sending insights to the cloud, using examples from oil rigs and manufacturing lines. Catterton also describes how the same behavioral profiling technology provides endpoint security by detecting anomalies in CPU and memory usage to prevent ransomware attacks and other threats.
Jeff Bier, founder of the Embedded Vision Alliance (now Edge AI and Vision Alliance), explains how embedded vision has evolved from expensive industrial inspection to ubiquitous consumer applications like automotive safety and smartphone face ID, driven by domain-specific processors and deep learning. Jens Gamperl, CEO of Sourceability, describes his company's hybrid distribution model, which uses tools like Source Engine to integrate real-time supply chain data directly into design workflows via APIs, bridging the gap between engineering and procurement.
The summary, chapters and topics are generated from the episode’s audio by this directory, not written by Embedded.fm. The episode itself belongs to its publisher.