Shiva Dhawan, Attentive.ai, and Creating takeoffs from PDFs

SPONSORED EPISODE

Shiva Dhawan, CEO & Co-Founder at Attentive.ai shares his journey from mechanical engineering to building an AI-based takeoff software for construction, and explains why manual takeoffs from PDFs remain one of the biggest bottlenecks in bidding workflows. The conversation dives into how AI can read construction drawings, the role of human verification in ensuring accuracy, and why PDFs, not BIM files, still dominate the bid phase in North America.

The discussion also covers:

• What construction takeoffs are and why they’re foundational to estimating
• How AI interprets drawings, symbols, and schedules from PDFs
• The challenges of inconsistent symbols and design standards
• Why human-in-the-loop QA is critical for reliable AI takeoffs
• Scaling from 50/50 human–AI effort toward a 90/10 model
• Why automation is essential as estimating labor remains scarce
• What’s next for AI in preconstruction, including estimates and bid filtering

This episode is especially relevant for estimators, preconstruction managers, contractors, and engineers interested in how AI can increase bid capacity without sacrificing accuracy.

For more about Sponsored Episodes, see our Sponsorship section. 

Transcript

Welcome And Guest Introduction

Roopinder 0:19

Hello and welcome to FoDES, the Future of Design and Engineering Software podcast. My name is Roopinder Tara. On the show, we will have guests that will discuss tools and technology that engineers will find interesting and useful. Our guest today is Shiva Dhavan, the co-founder and CEO of Attentive.ai, an Indian origin company that's looking to the North America market. Their product is Beam AI, which is an automated takeoff program.

Roopinder 0:48

Hi, hi. How are you doing?

Dhavan 0:49

I'm good. How are you? Good, thank you. Thanks for joining me. And welcome to the show. Thank you. It's a pleasure to see you. Where are you located right now? I'm located in Delhi, India.

Roopinder 0:58

Okay, very good. I'm from India as well. So

Dhavan 1:01

oh, nice.

Founder’s Path Into AI And AEC

Roopinder 1:02

You graduated with a mechanical engineering degree, correct? That's right. So what led you down this path towards infrastructure and AEC and construction?

From Insurance Imagery To Construction Takeoffs

Dhavan 2:07

I'll tell you a bit about my story. Yeah, I graduated from college, I wanted to start my own thing, but then my father said, Hey, you need to work somewhere, get some professional experience. So I joined a management consulting firm called AT. Uh, worked there for a year, and I was like, Hey, I can't work for anyone, I need to do my own thing. I was looking for different like ideas to work in. This was 2070. AI was just like getting started. I saw a few companies working in AI, and I was like, hey, this seems interesting. And what if I build a business in the AI space? And within AI, like there were multiple options to work on. So I was talking to a colleague who was a senior from my college, and he said, Hey, why don't you start with like doing services for any industry? Just do services, just build AI model, and through that process, maybe you'll come across a use case that seems interesting enough because sitting at home, you will definitely not find an idea, you know, like just break through because you haven't gotten that exposure. So I was like, Yeah, that that seems like a good plan. So I started uh like set up a company almost like a deep tech IT services company, worked with started working with different like companies and asking them, hey, are you looking, are you interested in building AI models? Specifically AI models on images, because I felt that a lot of people were dealing with text data, uh, where you know, like they were dealing with NLP natural language processing, and I was like, hey, everyone is doing this, can we deal with images? And then looked at a few use cases, started working with like PNC insurance companies in the US that seemed like a really lucrative use case where you know they had to go on the ground and like to ensure properties, they had to look at how much building square feet there is, etc. So I was like, hey, we can do this to aerial imagery. That's how it really started. Uh realized that you know, like this is like a treadmill, like just doing services, these PNC insurance companies, like this, what is the end game here? And then Serendipity like uh met a couple of landscape construction designers at a conference in San Diego, and they said that hey, what you guys are doing for insurance companies, if you can do something similar for like construction companies, extract like loop, like extract takeoffs from these blueprint designs that we get, because that's a hugely manual process. If you can automate that somehow, then that's that could be really cool. And that was sort of the like Eureka moment for me. This was 2019, just before COVID, and I was like, hey, this seems like a space which is not very crowded. I come from mechanical engineering, one of my co-founders was from Civil, so things came naturally to him as well. He was like, Hey, this is something I understand as well. Uh, and we can definitely do that. I saw that the space was huge, I saw that nobody smart, like, like not a lot of smart companies were building in this space, and that sort of culminated into us like going all in on the on the construction and uh field services space, essentially.

Roopinder 4:54

The attentive AI products and services were offered in India, right?

Dhavan 4:58

Yeah, we briefly for one year they were offered in India, but we realized hey, like North America would be a much better market uh for it, like primarily because of the cost arbitrage, but also because of the fact that the India, like India is a little more advanced in the sense that like CAD files are circulated everywhere, so it's much easier to do a takeoff. Whereas in the US and Canada and most of the West, they don't circulate CAD files before the project is won. So it's a PDF that is circulating, and that's why we have a big business because extracting that takeoff from a PDF is much more difficult than extracting a takeoff from a CAD file, like the material quantities, etc. versus a PDF. We were like, hey, let's focus here. This is a much bigger problem, and there are no CAD files, so it's much more difficult as well for our customers to do that.

Why North America And Why PDFs

Roopinder 5:47

But you had a a pretty good business going already in India, so it was it was easier for you to come to America and set up a business, probably because you had some already a business with revenue and that sort of thing. Is that what happened? Yeah, so you appealed for funding.

Dhavan 6:02

Our latest investor, uh, our series B investor is Insight Partners, which is a New York-based firm. But all of our previous investors are actually India and Singapore based. We got the funding from the Indian investors, the initiate funding, but our target market was the US uh from the very Well, you did pretty well.

Roopinder 6:20

I'm looking at the figures here. That's a lot of rupees. Are you one of the biggest venture capital successes in in our industry in in India at least? In the construction industry in India, yes.

Dhavan 6:30

Yes, I think we yeah, that'll be yeah, yeah.

Roopinder 6:32

You're sort of in the same industry anyway. So it'd be good to get you guys together. All right. So for the purposes of our audience, which is has some AEC and construction people in it, but largely design engineers who may not know what a takeoff is, or for them, takeoff is a different thing. Let's define takeoffs. So my very introductory definition, and then you can expand on it. Takeoff is really a list of materials, that list of material that needs to be bought for a construction project. Is that it, basically?

Dhavan 7:02

Yeah. So a construction takeoff, or sometimes also called a quantity takeoff, is the process of measuring and listing all materials, labor, and equipment quantities. So it's not just materials, they also list out what is the label needed and the equipment needed. Uh and this is basically based on drawings, plans, and specification that are part of the RFP. So takeoff is the fundamental step that is needed to be done in order to estimate how much this will cost. Without that, they can't really bid on a project, right? If they get that cost wrong, then they are basically underbidding or overbidding. So the takeoff is the scientific backup of that cost estimate that they give.

Roopinder 7:46

So it's part of a cost estimate, correct? It can't be the whole cost estimate because it doesn't involve labor. Or does your software actually determine labor costs? Right now, our software does not directly give them the labor cost. Some methods already that exist from the very manual methods of actually typing it into a spreadsheet to semi-automated methods. Bluebeam does take off. Yes, correct. Okay, okay. What drawing can it read? Can it read a Revit file? Does it read a Archicad drawing or all of them?

Defining Construction Takeoffs

Dhavan 8:26

It's possible to read a CAD file or a Revit file, but our focus is reading PDFs because that is what our customers send to us, like during the time of bidding when they're bidding for jobs. Most of our customers are commercial construction companies. And at the time of like they're bidding for a job when they get an RFP, they get the design in the form of a PDF only. So we need to process these PDFs essentially.

Roopinder 8:51

Do you have plans to files that are more data rich? Because I think Revit will store information, is it that could be useful for takeoffs?

Dhavan 8:59

Uh Revit can store information that'll be useful for doing the takeoff, but we don't have any of that information.

Roopinder 9:04

Okay. All right. Is that a future plan or is that you're just comfortable with doing PDFs?

Dhavan 9:09

I mean, like, if people start sending us uh Revit files, then we'll be happy to process that. I don't think that's going to happen. These people who are like it can be a general contractor or a developer who's basically putting out the bit, and they don't want to put an entire like CAD file out there in the open tens of construction companies are seeing that CAD file. They feel that it's their intellectual property, so they only send out the PDF. And that's what we also get then to process from the actors.

Roopinder 9:40

Reading a PDF and being able to determine a take up is somewhat like a miracle already. Does a PDF have to be neat, or can I just hand sketch something and have be able to interpret it?

Beam AI’s Focus And File Types

Dhavan 9:53

That's that's a good question. Like we get all kinds of PDFs, and uh like typically the fact that because it's most these are mostly commercial projects, yeah, the certain level of quality that we get uh and we've used those PDFs to train our models. We don't deal with residential projects if you just hand sketch something and send it to us, yeah, uh, which is more common on the residential side. Now I'm highly doubtful that they'll give you a very accurate output. One thing I want to clarify here, and you know, like just to make sure that you understand a little bit how our how our process works. So there is an AI running at the back that processes the PDF. We don't send the outputs directly back to the customer. There is a human in the loop that is sitting in India. There are 100 people sitting in India that are the part of our team that do the quality assurance on that AI output.

Roopinder 10:44

Did you say one 100? We have about 300 people. 300 people total, all of them in India or most of them in India? All of them in India. How many of them are involved with the scanning process?

Dhavan 10:54

All of them are involved in the scanning process. They do the quality assurance, and the AI generates the first level of the output, uh, which is like six, like does 50 to 60 percent of the work, and then quality assurance expert comes in and does the remaining work.

Roopinder 11:08

That's a lot of people. Are they architects by training or have you trained them yourself?

Dhavan 11:13

Most of them are fresh engineering grads from like not IITs, but like tier two, tier three colleges in India. Engineers, like mechanical engineers, civil engineers, and we've trained them internally. We call them estimators. That's smart.

Roopinder 11:27

I think if you hired all the grads to do your scanning, you'd quickly run out of your money. You said 40 to 50 percent is done by AI initially. Are you on a path to increase that AI involvement? Or are you happy with the balance?

Quality Control And Human In The Loop

Dhavan 11:42

No, we're not. Uh in fact, one thing uh that I wanted to highlight. So before we started doing construction takeoffs, we actually were doing maintenance takeoffs for the landscaping industry. We're analyzing aerial images and we were extracting like lawn measurements and sidewalk measurements. That was our first sort of product. And there we got to like a 90-10 ratio because that's a little more easy. It's fairly standard, like lawn is lawn and sidewalk is sidewalk everywhere. Construction takeoffs are a little different because it depends on the design and like how they design the whole thing. It's a little lawn standard, but we are not at all happy with the 50-50 number right now. Plan it to take it to 90-10 because we don't want the number of estimators scaling linearly as the revenue scales. We have a big team in India, AI team, about 15 engineers. We've hired a couple of people in the Bay Area who'll sort of augment that team, new funding. The idea is to get to like 90-10 in terms of the AI and the human effort. You're actually going more after the North America market now than the Indian market, you said. How's that going? How's the plans for market expansion? Going good so far. This year was pretty good. We grew about 130% uh this year in terms of revenue. Yeah, the goal is to grow at at least the same pace or double in revenue. Hopefully, we can get there.

Roopinder 13:00

You find yourself coming back and forth to US India quite a bit. But racking up those frequent flyer miles? That's a long, long trip. So what's your fly Air India?

Dhavan 13:11

Air India.

Roopinder 13:11

Oh, too bad. You must what's their highest status?

Dhavan 13:14

I have platinum right now.

Roopinder 13:15

Platinum, okay. Is that the highest, or is there like a ruby or something?

Dhavan 13:19

No, platinum is the highest.

Roopinder 13:20

I prefer not to fly them if at all possible. I don't think you have a choice. Your choice is usually governed by uh where most of the flights are, most of the direct flights. Do they have direct flights to America?

Dhavan 13:30

Uh yeah, they do. To New York into SF into Chicago.

Roopinder 13:35

Longest flight I ever went on was San Francisco to Delhi.

Dhavan 13:38

Yeah, that's 17, 16, 70 hours.

Roopinder 13:40

Yeah. How do you amuse yourself on a flight like that? Try to sleep or you probably work.

Dhavan 13:44

No, I I sleep pretty quickly. Like I try to sleep most of it, and then in the remaining time I'll watch a movie or something.

Roopinder 13:51

You watch American movies? Rewatch all this, all those movies. Song and dance.

Dhavan 13:56

I think Bollywood has really lost the plot in the terrible movies that are coming in Bollywood. So I don't like Bollywood anymore. Oh yeah, yeah.

Pushing Toward 90–10 Automation

Roopinder 14:04

I I know I've tried. I've tried. Have you seen uh Slum Dog Millionaire?

Dhavan 14:08

Yeah, I have.

Roopinder 14:08

They couldn't resist but having a dance scene in Slum Dog, but they saved it till the very end. Back to the takeoff. You said 40 to 50 percent right now. Where is the limits, BEAM.ai and the other product you have? What do they strain at? What's the challenge right now? Is it handwriting or is it interpretation of graphical elements? Where do you see the biggest improvement needed?

Dhavan 14:30

That's a good question. There are two parts to it. One is the draw, and one is like the schedule tables and everything. The tables are pretty standard, nobody gets them different or wrong. The drawing part where people use different symbols for the same thing, and that's where it gets tricky. Uh, for example, like in piping, there is a 90-degree elbow, which is like a fitting, and trained our model, identify 90 degree elbows, and like then we come across a plan where somebody has used a different symbol for the 90-degree elbow. And they've kept to the legend and they've said, Hey, this is a 90-degree elbow. But that's a new, completely new symbol for us. That's where it really gets tricky, where you know like people come up with their own symbols for pretty standard stuff that should be represented by a certain symbol, but now they're representing you using a new symbol. That's where the active learning component, where you know, like the models aren't like they're never done. Like they should be completely flexible to new data points that we see. And it shouldn't be like, hey, if somebody sent a new new PDF and you just learned that, we realize that that's like just a one-off case. It should be like, hey, you get a few samples, and then you're like, hey, this is a good enough sample size, and now we should train the model. But yeah, that's where it trips, where you know, people use different symbols in the drawing. Yeah, yeah.

Roopinder 15:45

I imagine the graph interpretation of graphical outlets would be the most difficult. In some cases, is there also a cultural problem? You know, there's like interpreting landscaping drawings in this country. You could have altogether different elements, right? Like India. I don't know if India even has lawns. In the US, it's like lawns are everywhere, mostly residential, but also there's like uh campuses and everything that have different landscaping. Do they have standards? I know the architecture does, but landscaping is seems to be like the uh sort of ungoverned. Is there landscaping standards at all?

Dhavan 16:20

No, no, there are no landscaping standards.

Roopinder 16:22

That's gotta be a really tough problem, then, huh?

Dhavan 16:25

It's not because like 90% of the people have sort of defined their own standard. This is their company or the largest landscaping company, about two billion dollars of revenue. It was called Brickman, and Brickman built these holy grail of like things of standards like labor rates, etc. And you know, design standards. So people, a lot of people sort of came from Brickman and they sort of followed that those standards at their own companies. From a cultural standpoint, just to answer the question that like we have trained our people on like US drawings from day one. Uh so they don't know actually Indian drawings, they haven't seen Indian drawings at all. They they only know US drawings. That's why just coming back to why we hire fresh fresh graduates to train them because they are not set in a particular mold. So we directly train them on the US drawings.

Roopinder 17:12

I'm not in the habit of doing this, but I did create some things that had lawn or turf, and uh there was a particular hatch pattern that they used that would signify that it was lawn area. You had to use your imagination, but if you did, but yeah, that might be that might be grass, a straight line with little clumps supposed to look like grass. You'd have to tell somebody that was grass half more than half the time. So yeah, I sympathize with good luck. What are your next big goals besides uh the expansion to US markets?

Where The Models Struggle

Dhavan 17:41

So the next big goal uh for us is that right now we just do takeoffs. Our goal is to become an end-to-end reconstruction workflow for these contractors, and that means doing other things apart from takeoffs, like estimates, like finding bids for them. These two are huge problems for them. Like takeoffs is the number one problem because it takes the most amount of time. But these two problems, which are the next two big problems for them, estimates is essentially once you get the takeoff, how can you connect the material aids from your catalog to that? That's also a little tedious because the simple VLOOKUP doesn't work there because the material price in the material catalog, the material name would be different, and in the design, so there is some amount of general-depth AI that is needed there to match that properly, and we are working on that. The second big problem that we are working on is just finding bids for them right now. They have logins to these different bid platforms where they get bids, but every day they get like thousands of options, and they need to sort these bids based on their expertise and based on what's their sweet spot in terms of what kind of bids go for. So we are developing an automated bid fill which uses AI to filter these bids and give them like a short list every day that these are the top five bids that you should bid on, uh, based on your company profile and your choice of projects and your like average budget range. Uh get more entrenched in their lives and you know, like make them like not have them just do take offs, but have a whole pre-construction workflow in BMAI that they can use.

Roopinder 19:11

It sounds as if the best that you're gonna use your funding to grow the company's technical abilities rather than just use it for the growth of staff or estimators. Quite a bit of it is for R&D.

Dhavan 19:24

Yes. That is correct. We can't scale the estimators linearly with revenue. AI is the secret source, a large part of the funding will go into the R&D team development. I hate the word R&D because whenever people say R&D, it's like, hey, this is a five-year effort. The word that we use use right now is like applied AI. Hey, can you apply the AI that's already out there and like fast track that transition from 50 50 human and AI work to 90 10? Yeah, 90 and 10% human.

Roopinder 19:54

Are is you're hiring IIT developers in India, and does IIT have a good program? Program for developers. The developers in the US are first of all very expensive. Second, very hard to get.

Dhavan 20:07

We are hiring on in India as well as the US less people in the US, like much more senior people in the US who can basically train some junior people in India who are really smart, but they don't have the necessary AI experience. We hired like 10 people from IIT Delhi and IIT in this, like for our software developers and our AI research engineers. The plan is to get that AI talent in the Bay Area. A few senior folks, not too many, but just very smart, very experienced senior folks, and then have a larger team in India that that can learn from these experienced people from in the Bay Area. That's how we are complementing the two.

Roopinder 20:44

I have to ask though, because in India, since we're both Indians are and both, I'm sure, both of us sets of parents wanted sons that were engineers or doctors. I'm sure made them very proud of you being an engineer. What did they say? Are you getting into software? Is that now a career for a person? Yeah.

Standards, Culture, And Training Estimators

Dhavan 21:02

In fact, uh they were really happy about it. Software is booming, and like for them, is like, hey, where are the jobs going? Like, where are people getting placed? 50% of the like batch gets placed in like software companies where the best packages are. So that's what they see, and they're like, hey, this is where you should go as well. And my parents were really proud 10 years back. You were sitting for placements, and now you're going there hiring people. So that was a pretty proud moment for them.

Roopinder 21:29

The parents are I think your employees are very happy. I see you had already had a an event. They call that when you give money to the liquidity event, liquidity events, right? You had a liquidity event already. Yeah, that's gonna make you very popular in your company.

Dhavan 21:42

I don't think it's good to be a popular leader. I you should be unpopular, but yeah, people were happy. Uh thanks to our questions because you know, like they did most of the liquidity as part of the fundraise, like they did a huge secondary, and like a lot of people wanted to sell part of their shares, so they were happy.

Roopinder 21:58

Congratulations on your success so far, and I wish you the even more success in the future. Hope we can have you back next year when you'll have your 90 10 ratio. All right, Shima, very nice to meet you. And uh what time is it now at India? I must be late at night. Thank you very much.

Dhavan 22:13

Absolutely pretty used to it working on cars, and like this is not too late actually. It's 11 10 p.m. So it's not too late as well, so not a problem at all.

Roopinder 22:21

Although I appreciate it. Very nice to meet you. Okay, all right for joining me and uh get some sleep. Thank you for listening to FoDES, the Future of Design and Engineering Software show, brought to you by ENGtechnica. I hope you have learned of a new application or technology that will help you with your job. If you have an application you think would be of interest to other engineers, please let me know by emailing me at roopinder at engtechnica.com or message me on LinkedIn.