Introduction

My name is Maharam. I'm the founder of ClaimGen. Today I'm going to talk to you about what we learned using LLMs while we were developing ClaimGen.

SR&ED ("Shred") and the ClaimGen context

So let's start first with Shred. I saw that a few people know about Shred here, but for people who don't know, Shred is R &D tax credit program of Canada. Every year Canadian government gives around 4 billion to eligible businesses businesses, and we help Canadian businesses to access this funding.

What claim preparation requires

So one of the main parts of our work is to understand what a company did technically and detect what qualifies for Shred, and then write a technical narrative based on our findings.

From "AI can write it" to automating the whole workflow

So when the LLMs became popular, I thought that we can just use AI to generate these writing so we don't need a technical consultant anymore. So we started doing that and we realized that actually writing is easy part and it's not the most valuable part that as we saw it will be.

Because generating writing with AI as all of you have done probably before is easy but what happens before that is actually what mattered for us.

So we started to look at every stage of the workflow and understand okay what should we do to automate this manual parts of the workflow rather than just focusing on their writing so let's first focus on what we have done and then I'm gonna go to the lessons that we have learned

Process improvements and timeline reduction

through our development process the first thing is that traditionally a shred claim preparation takes two to three months so it has technical calls interviews, project identification stage, writing a report, follow -ups, document collections and more and more.

So a client can spend up to 40 hours inside of this process. By the help of the LLMs we were able to reduce this under four hours for client time and under two weeks for whole preparation process.

But of course it didn't happen overnight, it didn't happen just like fixing the writing part of it, there was multiple things that we had to do before that.

Automating the technical interview with an AI interviewer ("Genie")

And let's focus on one of them, which was automating the technical interview process of Shred application. So instead of, again, let me give you the example of how it goes first.

So traditionally, a Shred consultant have a call with a client, and they start to talk about something that happened six, seven months ago, because Shred is after the fiscal year end.

So when they started to talk about that, the client usually gets uncomfortable because they don't remember what happened six, seven months ago exactly, and they have to answer consultants' questions on spot.

To take out this awkwardness, we decided to build an AI interviewer which client can use in their convenience and can feel comfortable about it. And it became one of the most valuable things that we built

because the clients could just go to our platform, platform, use our AI and do it whenever they want rather than scheduling a call with consultant. But the main thing here is that it also became very valuable for us as well.

Lesson 1: Better inputs drive better outputs

Because the inputs that we were getting from Genie, our AI interviewer, was very valuable for the next stage of our whole system.

And it made us understand that, okay, once you have the better input, you can get the better output.

So, the quality of everything downstream depends on the quality of information you collect upstream.

Lesson 2: Your AI has to push back for specifics

The next lesson we had was that your AI has to push back. It cannot accept everything that client gives to it. It happened during first versions of the Genie.

The client will give an answer to the Genie and Genie will just accept it as it is and just move to the next question, say thank you to the client. But when we check the interview, we'll see that, okay, we actually didn't get the technical details that we needed from this interview.

So what we did is that we started to show Genie, okay, how the good answer looks like. For example, client would say we improved performance of our system, but that wasn't enough for our industry for short application process.

Because we need to understand why did you do that? What approach did you take? and multiple things over over this basically so the main things that we

learned over here is that what matters is you need to make a I understand how the good answer looks like rather than giving it a context and expecting it's gonna do it by itself okay now let's move the next lesson and the next lesson

Lesson 3: When AI reads AI output, add audit and verification layers

is I think it's very common right now a ice your AI is reading other AI output But right now, because before the client documentation was mostly written by other people, but right now, what does client do?

So the clients give their raw data to cloud or ChatGPT and ask it to summarize it for shared application process. So the cloud and ChatGPT, which doesn't have any knowledge of Shred, just summarizes it and uses some fluffy words, which our AI thinks is Shred eligible. So, when our AI gets that and thinks that it's eligible, it makes a problem for us.

Because the human goes check it and understands that, okay, actually, the result of this wasn't eligible.

What we did to stop that, we built multiple audit layers. And I think this is very essential for now for every company to do this kind of thing.

So we have audit layers, we have agents checking each other's work to confirm, okay, it's actually accurate or not. and we have a verification throughout the whole process for our AI tools.

Okay. Basically, don't accept anything our AIs give to you.

And the next lesson we

Lesson 4: Prevent detail loss in technical writing

have, this is the last stage of our tool, so once we have everything ready, Gini did the interview, we got the information from the client, our audit layers confirmed that okay, it's accurate, it's shared eligible, we have now technical writing stage.

How do we generate the report? We put all the technical details to our writing tool and we expect it to generate us the final result.

But the problem here is that LLM tends to compress the details. They like to just generalize everything and for our industry, for Shred, that's one of the main problems because in Shred you need as much as detail as possible.

So for this we had to make our AI understand that you have to interpret everything and you had to understand, okay, which information is essential and needs to be in the claim rather than just compressing it and making a general assessment from that.

So that's pretty much it. That's our experience using LLMs while building claim gen.

Key takeaway: A hybrid AI + human validation approach

And one thing that I want to mention here is that the main understanding that I have, especially in professional services, is that you need to have a hybrid approach because just letting AI to do the work is not enough. You need to have humans to validate everything that comes out of the AI, otherwise it's garbage in, garbage out.

So that's pretty much it. Thank you for listening. If you have any questions, feel free to ask.

Any questions on how to get free money?

Today we didn't talk that much about Shred because I didn't want to bore you about it. we mostly talked about how we use it in our development process

Q&A: Building agentic systems with safeguards

Hi, my name is Raisa, thank you so much that was a really great presentation especially that last point you mentioned

regarding how you really have to make sure that there's a human validating those steps that really resonated with me

because I'm building in the neurotech space and you know, especially with anything that involves healthcare or life sciences

you really need to make sure that you're hitting those regulatory concerns concerns.

So I'm wondering when you're building an agentic AI like this, how do you make sure that you're actually addressing those concerns as you're building?

So the main part is that you need to understand where the human has to be in the loop for the process. So, okay, you made your agents, but okay, in which process you're entering the human to the process.

For us, that was the essential part because if the human comes only at the end, just to check the final result, you're to have human in multiple layers rather than in the end so that's what i would suggest just to

follow up thank you for that answer but from my perspective i don't know a lot about claim gen but from what you just shared it seems like it's more of a tool that kind of cleans up the workflows and then has some agentic components how did you come up with this model so uh again like the first

i saw that the writing part will be the enough but then when i try to simulate a shred consultant live to an AI, I realized there are actually multiple steps.

And that's how I came up because I have an expertise in Shred and I know everything about the Shred. So I know which processes can be automated, which processes should be left to the human.

So basically, once you know the industries that you're working in, I think you have a little understanding of, okay, how can I move on and how can I automate the process? Thank you.

Q&A: Tech stack and controlling writing style vs. generalization

Hello. Thank you very much for your presentation. I have a question about the tech stack that you used for GenClaim and another specific question about how you made your LLM able to generalize and include detail at the same time in other contexts.

Okay. Let's start from the second one. We don't actually generalize. We don't allow our AI to generalize anything the same, because we need as much as detail as possible in our Shred industry.

And coming to the first question, what tech stack we use, we have API calls from OpenAI, Cloud, Anthropic, I'm not a technical person behind it, I'm the business guy, but these are the general tech stacks that we're using.

If you're looking at the LMS files, we don't have our local host yet, we are using others, others, but as a company grows, as a startup grows, we are thinking to make our own local LLM rather than relying on the anthropic or open AI ones.

So just a final follow -up question. So you said that you don't generalize. Are you using it mostly to extract fields to fill forms?

So what we do is that we give the AI examples of how it should be written, how detail -rich it should be, and we expect the AI,

we see that our AI is actually copying the writing styles that we have from the examples and just making the similar assumptions based on those examples.

So, yes, we don't allow it to generalize because, again, a generic writing report will not survive to in Shred Audit. If you have dealt with Shred or not, I don't know this,

but in Shred Audit, usually, CRA wants to see the specifics of your claim rather than just a generic assumption. Thank you very much.

yeah so my question is um you were talking about having the AI not make generalizations which AI is really good at doing and not just putting human at the end for quality assurance so let's could you give like an example let's say you're writing a cover letter maybe that's a basic thing that I could think of right now how would you put the human in that process not just QA at the end

Okay, let's start from our own process first. As I mentioned, we have multiple stages, like interviews, like document collection. After each stage, our human checks what our AI has done. In document collection process, okay, did it identify the shared eligible work accurately or not? After that interview, okay, did it ask the questions that we wanted to ask or not? So every step, you can include the human in the loop.

And coming to the writing cover letter, that's a very interesting question, to be honest. But what I would think is that when you write a cover letter, you start from, okay, here is my background, here is my education, here is my work history. So let's say AI writes your background, check it. Let's say AI writes your education, check it. AI writes your work history, you got that. You check in each step rather than just checking the final result of cover letter.

Q&A: Governance, controls, and who stays in the loop

I think you touched upon quite a bit of it, but I just want to know more about when you are talking about the need for regulation and scrutinizing AI activities. So you touched upon something, peer reviews, audits, and seeing how the steps are going, but what are the kind of, like, what specific controls you have in place?

Like are there specific bodies within the organization? I don't know the strength of your organization because I came a bit later, but whatever team you have, do you have dedicated people and how do you tend to maintain and optimize those controls over time?

Yes, we have dedicated experts, shared experts who have been in the industry for a long time to understand when the AI says something is shared eligible, but it's not. So it's essential for you to have the people from the industry first in the company to validate the answers of the AI.

And also from the technical perspective, as I mentioned, we have built audit layers, agents that check each other's work. And that's coming from our technical person's capability, which is our CTO.

Thank you. Thank you so much.

One more question. Thank you.

My name is Khan.

I wanted to ask, because you mentioned that there has to be a human in the loop in most parts. You mentioned nearly every single part. And how does that happen? Is it usually the user that has these different things that they come in the loop or is it you that is making these decisions because that can get very nauseous at some point, no?

True, true, true. But you can't have the human, the client in the loop itself because the client doesn't know the shred. You need to have the shred expert, the expert of any industry, it doesn't matter. In our case, it's shred. You need to have the expert of the industry to be in the loop and understand what's the result coming out of the AI. And it's usually saving more time rather than that person going to collecting the data interviewing the client and doing all the processes itself rather than just checking the Result end of each step.

Okay. Thank you.

Okay, so one one last one we have we can take one more Okay, Maram is a very nice guy very patient guy

How how big is your data data in your database, how vast is your traffic with respect to the customer base and so on, and how have you automated the data validation and things like that?

Yeah, I think we touched the data validation stage. We have multiple layers, agentic layers and also human layers in that, and when it comes to how big is our data, to be honest, I'm not going to be able to say that.

do that. Even if I knew, I wouldn't say you that, to be honest, because it's a confidential company data, but yeah, that's all I can say.

Nice, but not that, I mean, you can't... I'm not going to share all the details of the company, come on.

Okay, so, okay, one last question. I said that in the last question also, but this has to be the last, okay?

The food is here. I cannot keep people, Maram, between the food and them. Yeah. You want to say that?

Okay, my name is Ethel. I just want to make a contribution and support where he's coming from, like a non -technical background. I'm a technical background.

So with large language models, it's not every model that is going to generalize. back and if you have control of the back and you're able to set parameters of how much it can give you feedback so when he's mentioning about us having a human being having to assess whether whatever we have retrieved from a large language model we okay the best way to say it is that

let's say we've got four layers right of our of our software development that is happening When talking about a cover page, like the example that was given, we have to know what is critical, what is important, what are the security details, so we can program our large web model, give it keywords, what is it going to search to avoid hallucinating.

So when we also talk about data, like how much data is it being trained on, like we we take example of ChatGPT, it's trained on large parameters, like more than billions, right? So it's going to be hallucinating a lot, which is why when you prompt ChatGPT it can say anything, but now, not from the back end,

you can also customize the front end of ChatGPT as a user. So you can go to the settings, you ask your, you know when you prompt ChatGPT, it can give you answers that have got hyphens, It can bring emojis, you can set your settings from the front end as a user with no coding.

You give it an instruction. I don't want my responses to have these hyphens. I don't want my judgments to be biased. You can make it stop hallucinating and generalizing the output you're getting from it. I hope that helps.

Thank you for volunteering and answering. Great explanation, to be honest. Thank you so much.

Amazing. Amazing.

You know, how many job seekers in the room? Job seekers, we always have some who are looking at a potential role.

Closing remarks and hiring plans

Maram, not to put you on the spot, but you know, you guys are building a product. But what's the next? Is there an open job opening right now?

We don't have an open job right now, but we will have very soon. So feel free to reach out to us through the year. And we are going to keep your information. When the time comes, we are going to reach out to you.

Okay, so what is the role that you look maybe in six months down the lane? The main roles that we're going to look right now is marketing person, content writer, basically. If someone has experience with us, any SEO, geo person, that can be one thing that we are looking for. And the next person that we're going to look for is we're going to actually train someone to become a technical consultant on top of our AI. So if you want, just reach out to us.

lot of technical consultants in the house so i guess they will reach out to you whenever the time is right so thank you for the presentation thanks everyone hope you enjoyed it too round applause