Implementing an end-to-end AI product requires a variety of skill sets, especially if the product has to take on a physical form. IOT engineering, building the AI models, data engineering, and compute optimization are some of the technical challenges involved. At the same time, designing for success needs to take on the tougher challenges of working with the eco system and at times, even shaping it. An AI startup, Vulcan.ai has dealt with every one of these issues and can share with us the reasons why such ambitious projects might fail.

Manik Bhandari was the ASEAN Leader for Analytics at EY, based in Singapore. He has been a Management Consultant with Accenture and EY and was part of the founding leadership teams. He has worked with the Singapore Government, especially through EDB to set up multiple Digital and Analytics Hubs and Innovation Centers which has contributed more than $100 million to the GDP. He also serves as the Vice President of the Business Analytics Chapter of the Singapore Computer Society and actively contributes through his business network to build local capabilities in the Singapore eco-system.

Kamal Mannar is an accomplished AI leader with experience in innovating and bringing data driven solutions to industries in the public and private sectors. As head of AI in Vulcan-AI he leads a team in developing innovative products that leverage the latest in AI, IoT and data sources like remote sensing to create breakthrough capabilities. He has led development of innovative capabilities like deep learning, reinforcement learning, computer vision and optimization. He has multiple patents and publications related to application of AI in healthcare, prognostics, health management of assets, smart grids, and precision agriculture.

[00:04:52] There's a Venn diagram that said an ideal AI person has software engineering skills, AI skills, business acumen. It's not possible to have all the skills in one person.

[00:15:07] You can take a few moonshots when you know that your cashflow is okay.

[00:16:40] We want to embed our AI algorithms on a very small footprint.

[00:27:40] You have to build your model from scratch because the requirement is not just accuracy. Is how much power is it consuming? And how long does it take to run on an edge device?

[00:28:35] The idea is you're not building the hardware for the AI, you are custom purposing the AI to the hardware.

[00:29:11] How many flops, which is floating point operations per second, is each model consuming, how much power is it consuming? And therefore, how do I optimize it?

[00:33:54] Basically we were tracking, this is our target cents per hectare.

[00:37:04] Every piece of code is always profiled, not just in terms of the accuracy, but in terms of run time, in terms of logs, for consumption and so on.

[00:41:10] ML engineers: these they're more, pure software engineers and especially because the data we handle is quite huge, it's streaming data, it's images.

[01:10:33] Companies will start differentiating their AI solutions based on, a green mark platinum or something based on the energy consumption.

Vulcan AI’s website: https://www.vulcan-ai.com/