Live data from Hacker News

Launch HN: VergeSense (YC S17) – AI-Powered Sensors for Building Management

news.ycombinator.com

11–16 of 16 posts

Re: Launch HN: VergeSense (YC S17) – AI-Powered Sensors for Building Management

#11
post #4

Have you looked into IoT systems control? Things like temp monitoring feeding into A/C draw, electricity off when people leave etc? Utilization of space is,one consideration, but a major other factor is maintenance, which comes down to optimizing running costs and minimizing wear through preventative maintenance and proactive design - IoT has a lot of potential in that space.

Great question - we believe there are a lot of potential add-on modules, especially around building control. One that is gaining a lot of interest recently is using people-counting data to more precisely control HVAC systems (most systems today rely on simple motion sensing for control). Modulating heating / cooling based on the exact count can help cut energy consumption, sometimes by as much as 30% for commercial b…

I was talking with a maintenance supervisor for a large facility and he was saying that they were able to modify the amount of make-up outside air cycled into a facility based on occupation (think O2 depletion, lol). Is that something you folks have run into? Seems like if you could avoid exhausting a couple (hundred?) thousand cubic feet of cooled air per hour, you could save a good bit of money.

Re: Launch HN: VergeSense (YC S17) – AI-Powered Sensors for Building Management

#12
How do you defend the statement "Each sensor creates a sphere of intelligence and the more data they collect, the smarter they get."

Do you mean each device gets smarter individually because the specific device learned more about the specific space? Or that there is some kind of supervised learning component where you would adjust the algorithm/model over time for every device.

Re: Launch HN: VergeSense (YC S17) – AI-Powered Sensors for Building Management

#14

How do you train the models in the cloud, run servers yourself, or use some service?

Currently we train on AWS EC2 instances. In the very beginning I was training on the GTX 980Ti in my desktop, which actually performed way better than I expected. Models trained faster on that machine than on p2.xlarge instances on EC2. But the advantage of training multiple models simultaneously, and using multi-gpu machines made the switch to EC2 worth it.

Re: Launch HN: VergeSense (YC S17) – AI-Powered Sensors for Building Management

#15
post #12

How do you defend the statement "Each sensor creates a sphere of intelligence and the more data they collect, the smarter they get." Do you mean each device gets smarter individually because the specific device learned more about the specific space? Or that there is some kind of supervised learning component where you would adjust the algorithm/model over time for every device.

At a local-level, each sensor builds a background model, which we diff against & combine w/ inference outputs for detections (background modeling helps reduce our false-positive rate). At a global level, we continuously push new pre-trained models over-the-air. These are built using 3rd party data sources (so not sourced from the sensors themselves).

Re: Launch HN: VergeSense (YC S17) – AI-Powered Sensors for Building Management

#16

It’s a super interesting space: we’ve been working on a similar concept here at Shepherd (Shprd.com) for a couple of years. We use existing SCADA & BMS embedded sensors as well as industrial standard retrofit sensors to send data to our cloud analytics platform. Uptake is strong, as you say, because facilities management can benefit a lot from condition-based monitoring enhanced with ML. Good luck - reach out if you…

Sent you a note!
Post reply on HN