Embedding vision AI in low-cost devices.

We are developing an AI-powered operating system, VisionOS, for low-cost security cameras that will enable them to detect objects such as people, vehicles, or animals in a live video.

Today, these security cameras lack the ability to detect objects in a live video feed as object detection is computationally expensive and the low powered hardware of a security camera makes it impossible to do so.

With our technology, the security camera will be able to run object detection on the device itself and in real time. Our technology is up to 62x faster than the state of the art methods.

New Delhi, India / New Delhi, India

Before working on this idea we were building a chatbot that could answer questions by searching inside videos like: “When was this taught in this lecture?” and it would pinpoint the location in the video when it happened. The demo is available on the website:

https://learnvenue.com/#demos – “Spoken Phrases demo”

We made progress with the chatbot and reached out to the CMO of India’s largest online course provider. He looped in his fellow CXOs and VPs in a video call with us to understand the product. We then had to back out as they literally asked us to log into our servers and show how it’s done.

After this debacle, we reached out to other online course providers and they were all very keen on using this technology but they couldn’t find a prominent use case for it. We then extended this idea to searching real-world objects inside videos such as people, vehicles, or animals. While working on this problem we realized that our computers were painfully slow to analyze videos for objects – this was the premise of this startup.

—- Saurabh —-
I have got a patent on the world’s cheapest, $100, Braille printer. I hacked an XY plotter to work as a Braille Printer.

I’ve also pitched a product on a nationally televised show named “Pulse The Venture” on CNN News 18. It was a chatbot that could answer scholastic questions.

I’ve built a startup with 4 fellow students in college and ran it for a year. It was a platform that connected students with vocational training institutes.

—- Nishchal —-
NGOs to impart academic and moral education to 100’s of homeless and underprivileged children in Delhi for 4 years.

I got admitted to the University of Edinburgh, one of the top 20 universities in the world, on the basis of my academic performance and projects in machine learning.

I was awarded Scholarship for Higher Education by CBSE, Govt. of India for scoring top 1% in science stream nationally.

—- Saurabh —-
Early in 2016, I participated in an event called Startup Weekend. The rules were: participants had to pitch their ideas and they were put to vote. The ideas/people that received the most votes had to form a team with the other participants and work on it.

I pitched my idea but it didn’t make it through the voting stage. I went to the organizers and told them that I want to work on my idea but they just reiterated the rules to me. I persisted and finally, the organizers challenged me that if I can make a team of at least 4 I can work on it.

I took the challenge and started approaching the participants, cajoled them, and successfully built a team of 6 – which was the largest. We became the runners-up of the event and won $500 as well.

—- Nishchal —-
I conducted several workshops during my undergraduate years and faced a trivial problem; we needed separate permission from the higher authorities to conduct them and each permission was valid for 2 months only.

Once I needed the permission for a whole year. I figured out a gap between how the higher and lower levels of management operated in my college: once approved at a higher level, the lower management never questions it. I intentionally wrote the permission letter highlighting the starting date of the workshop on the first page and added a paragraph on the second page stating the requirement for the whole year. Generally, the higher authorities are short on time and skim the first page and take decisions accordingly.

My permission application got approved by the higher management. I went to the lower management and highlighted the part of the requirement for the whole year. They approved it without questioning the duration.

We have known each other for almost 7 months. We first came in contact in July ’17 when Nishchal was going to return to India after completing his masters in AI. He was looking for opportunities in startups and discovered this. We discussed and assessed each other for over a month before deciding to work on the opportunity of video intelligence.

Nishchal came back to India in October ‘17 and since then we have been meeting and working together. Within this time period, we have completed a product that can search inside videos and pivoted it to achieving on device intelligence.

It will take 2 more months create a functional OS that can be integrated into the security cameras. However, the core technology that enables AI models to run on low-cost devices is working today.

We have been working for almost 4 months together on this. The startup is, however, 2 years old and I along with other team members – most of whom have left – have been working on it before I boarded Nishchal as a Co-Founder. The idea has been pivoted 3 times in those 2 years. Nishchal and I both work full time on this startup and do freelancing for our expenses.

We have demoed the prototype to an IP camera manufacturer in the US and to a local factory owner. The factory owner made us an offer of $250 to install it.

NA

NA

We stumbled upon this opportunity while we were building a chatbot that answers queries by searching inside videos. We realized that our AI models ran very slowly on our computers. It took around 1.5 Hrs to analyze a 1-minute video. So we decided to venture into research and figure out a way to make it run faster.

Nishchal has core expertise in AI and I have expertise in UX. Plus I possess horizontal skills ranging from deep learning, UI development, managing servers, and backend services. We also are in touch with industry experts and through them, we are understanding about the security/surveillance space.

We have got in touch with an IP camera manufacturer based out of US, and have discussed the possibility of integrating the OS into their hardware. They are pretty keen on our technology and we are in verbal talks with them.

To understand whether the end consumer will benefit from AI-powered surveillance, we got in touch with a local factory that was installing security cameras. We gave them a demo on a raspberry PI that detected people as they appeared in the frame without using the internet. He offered us $250/camera to install the system.

We are making new AI models that enable low-cost devices to analyze videos on the device itself, thus eliminating the need for relying on Google and Amazon cloud intelligence or investing money in costly computer hardware. Our AI models can run 62x faster than the current state of the art methods.

In the security space, the video is analyzed only post an incident. Which today is done by humans. ‘Netatmo Presence’ is a security camera that can classify objects such as people, car, and animals in a video.

Recently, Amazon announced ‘Deep Lens’ video camera which is also able to “classify” live video feed and they are targeting it at developers. It is a step in the right direction.

However, we must mention that image classification is not the same as object detection and is at least 20x less computationally expensive.

Plus, there are IP cameras in the market from Nest, Blink, Hikvision that can detect motion and then trigger a notification to the owner. However, it cannot differentiate between a motion caused by innocuous objects, such as pets, and intruders.

Our direct competitor is XNOR.AI. It is a startup based out of Netherlands. We both share similar technology and vision to make the device intelligent rather than the cloud. Google and Amazon Cloud video intelligence also pose a threat indirectly. There is always a chance that Google tries to venture into on-device intelligence. However, we believe that it will be against their established infrastructure and services around cloud intelligence. But we do fear this the most.

OEMs do realize that using cloud or an on-premise server is not the best way to analyze video content from all the security cameras that are sold by them. Firstly, there is a perpetual cost associated with using it and secondly, there is a limit to the number of videos streams that can be analyzed at the same time with the limited computational power.

For consumers, if the security cameras were to use the cloud, then the bandwidth required to upload the camera feed would be over 40 GB/day – just to figure out if there were unwanted people or vehicles in the video. Worst if the network breaks down leaving them vulnerable.

Making the security camera do the job of identifying the objects is the best way forward. The camera must be made intelligent in itself. And thanks to our technology, the low-cost devices can now analyze the video in real time without the need for depending on the cloud.

We will be licensing the VisionOS to the OEMs such as Bosch, Panasonic.

We charge $25/ device.

Security cameras sold in the World in ‘16: 52 Million units; 17% CAGR

TAM : 52 Mn x 25$ = $1.3 Bn

SAM (IP cameras): 38 Mn x 25$ = $950 Mn

https://www.sdmmag.com/articles/92407-rise-of-surveillance-camera-installed-base-slows

Since we are a B2B startup getting the right connection to the OEMs will get us there.

NA

Nishchal writes the code primarily for the on-device intelligence technology. We together research on it. Servers and designs are handled by me. We have one more team member (non-founder) who is developing the OS.

People have worked and left the previous pivots of the startups – non-founders. Also, they were not legally our employees.

Our technology can also be used in:

Developing an SDK:
The SDK can be used by mobile developers to harness the power of the mobile device to run computer vision applications.

Augmented Reality:
To detect objects in a live video stream. Think of Google Glass with added object tracking to achieve near ‘Iron Man’ capabilities. Drone OS: Drones are essentially low powered computers and our technology can provide them with much needed visual intelligence data for navigation, object tracking, or object avoidance.

Reducing the number of servers:
If our technology were to be deployed in servers, it will require less number of them to run AI models such as image classification, speech recognition, or video intelligence.

Vision for the blind people:
Once integrated into Google glass type of devices it will deliver much more useful information about the environment to the blind person than current ‘smart’ cane that only tells if ‘something’ is ahead of them.

Founderspeak – /faʊndə-spik/, noun

Speech used by founders to get the attention of people and investors at large with an ulterior motive.

“I’m not looking for money, just your feedback.”
Translation: I need your money.

“We have 100s of organic sign-ups.”
Translation: We forced 100s of people to sign up.

I’ve been following YC since 2015. Last year in October, YC held office hours in New Delhi, India. We got selected for it and spoke to Anu Hariharan about the stuff we are building. Those 10 minutes were enough to question ourselves about the startup. She convinced us as to why technology companies proliferate in the Valley. We decided at that time to apply for the next batch (S18).

Recently one of our mentors has advised us to look for acceleration opportunities outside of India as it will be hard for us to find the ecosystem that supports technology startups in India.

Plus, our target market is mostly outside of India and being in the valley will make sense.

Last year of my college (2015). Been following it since then.