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

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

36 pointsdpryan15 comments

Hello HN! This is Dan and Kelby (tripleplay369), the founders of VergeSense (http://www.vergesense.com). We're building an AI-powered facility management platform that helps companies use their buildings more efficiently. The cost of real estate is typically the #2 cost center for any company (after people), but most companies don't have a good way of measuring how their building is being used. Our product solves this by identifying wasted areas and recommending more productive uses for that space (e.g. turning unused offices into conference rooms or employee lounge areas).

The core of our offering is a discrete sensor that leverages multiple inputs (primarily an imaging sensor + PIR-based motion sensing), which feed into a neural network model that executes inference directly on the device. This allows us to do powerful processing on inexpensive hardware.

Our machine-learning stack is built around Tensorflow, which we use in two ways: 1) for inference (we embed Tensorflow directly on a Raspberry Pi), and 2) training new models in the cloud. New models can be pushed remotely to the devices over-the-air to make the sensors “smarter”.

While our sensors are currently trained to count people, our vision is to evolve into a 100% passive "super-sensor" that can be configured to detect thousands of different types of events. Examples that we've explored include things like detecting falls (e.g. during an emergency), counting assets (equipment, furniture, cars), and monitoring equipment usage (for preventative maintenance).

We're happy to chat and would love to hear your thoughts. Some things we've worked on that might be interesting to discuss: rapid-prototyping for hardware (Raspberry Pis +ESP8266), machine-learning, computer-vision, building automation, BLE, B2B sales, keeping sane while drawing bounding boxes, or anything else that comes to mind!

We look forward to your feedback!

Dan + Kelby

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 want to chat,

Will

Sent you a note!

TechCrunch previously wrote about VergeSense. See my post:

https://news.ycombinator.com/item?id=14947275

Tried to change the title of that post to:

VergeSense’s (YC S17) AI sensing hardware wants to reduce the usage of office space

but HN didn't let me add (YC S17)

Hi!

- What about privacy, is filming workplaces in high resolution ok with customers, their employees, the law and unions?

- Your FAQ states that you are selling the whole package for a yearly fee. Isn't that quite a risk when the customer is using mobile data as connectivity and having devices in the field that can and will fail and have to be replaced? Do you pay then a contractor to replace a single hardware node at your customers location?

- Have you looked at warehouses as customers? I suppose real estate is their #1 cost center :)

- One of the benefits of our system - as opposed to other more traditional video-based approaches - is that we never send any raw data off our devices. We have a light-weight neural-net model (about ~10MB) that runs directly on the devices, and only reports back on detected events (so things like “person detected”, “door-entry passed”), etc. This also has a side-benefit in that our devices can operate on low-bandwidth networks (and makes it economical to backhaul detected event-data over a cellular network).

- We include a gateway device with our product, and if anything goes wrong (sensor or gateway goes offline), we cover this as part of our service contract.

- Warehouses are another potential vertical, provided we have access to training data to train up our models. For example, if someone wanted to “count” things like boxes / forklifts / etc, our sensors can be configured to detect them.

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 buildings.

ARPA-E (Advanced Research Projects Agency) recently put out a proposal for such a system - you can read more here if you're interested:

https://arpa-e-foa.energy.gov/

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.

I did some work with a simple energy management system for a large retailer that had 4800 stores. The amount of money that can be saved with simple things like delaying HVAC start at open based on ambient temps, or using sunrise calculations to know when to turn on/off lighting is pretty ridiculous.

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).

Could this be used in retail store spaces to get footfall analyticsc?

Also, what this the average area the sensors cover

Retail analytics are a potential use-case, but we've decided to focus on space-optimization for buildings because we believe it's an underserved market relative to the opportunity. There are a lot of established companies doing footfall analytics for retail (using things like WiFi, BLE, door-counters, thermal images, analytics on video surveillance data, etc.).

As for sensor coverage, we cover about 1k sqft per sensor (it'll vary a bit depending on mounting height - higher mounting equates to a wider area of coverage)

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.

Reply on news.ycombinator.com