Donald Allison - The Big D, the brains behind KnowledgeFlow

Six legs on the pantomime horse this week. This time the third stool goes to the man who started it all, the on, the only Donald Allison, CTO and architect of KnowledgeFlow. 

The origin story is better than any of us remembered. In June 2022 Neil was getting ready to retire. Donald said don't do it because I've built this exciting thing with AI. In August, Kieron and Neil went to the biggest AI conference in the world in Vegas, got hideously sunburned, avoided the gaming tables, and had a minor incident in the Bellagio bar that remains permanently classified. They came back, went to see Donald, and a month later Leading AI existed. Donald had built a private ChatGPT before RAG had a name. Everyone he showed it to said the same thing: that's magic. Nobody believed it could do what it did.

Three years on, this is the conversation about what's actually under the bonnet. It's the most technical episode we've done, but hopefully everyone will learn something from it.

Why KnowledgeFlow lives in your Azure tenancy. Donald's answer is the cleanest argument for it we've heard: your security is already there. Role-based access, the lot. So why would you break that barrier and drag your board minutes, CVs and medical notes somewhere else — just to make the software easier to write? Better to write software that complements your security posture than one that dismantles it.

The safeguarding filter that catches the question, not the answer. Edge devices filter what comes back. But if a student asks how to build a bomb, the LLM refuses — so the edge device sees nothing, and nobody ever knows the question was asked. Donald built it the other way round. Catch it before the LLM ever sees it.

The browser extension nobody talks about. Use Azure or OpenAI APIs directly and internet retrieval goes out through their connection — quietly bypassing every firewall and filter you've got at the edge of your network. KnowledgeFlow routes it through the user's own browser, own gateway, own firewall. Kieron's summary: "So we're even more secure than Azure."

The hardest thing he's ever built. The hardest thing to do in AI isn't unstructured date (like 700-page legal packs, or 1,00 PDFs). It's structured date like spreadsheets. A SQL database with 50,000 rows is harder than a library, because the differences between line 1 and line 427 are minuscule — and LLMs are built for unstructured data. As Kieron puts it: it knows the difference between a kettle and a car, but 10 and 11 look basically identical.

The BidWriter argument. A customer handed it back this week. Kieron asks the uncomfortable question: should we retire some of the early tools? Neil defends it hard — 60% of a bid in two hours versus two weeks — and reminds everyone of the 190-question spreadsheet he cut and pasted for two days before Donald said "I can fix that" and built the button. Donald's verdict, and the line of the episode: give me a trowel and a pile of bricks and ask me to build a house — imagine what that's going to look like. Plenty of people can prompt. Most can't. That's who the tool is for.

And the thing Kieron caught him building. Automated downtime alerts started pinging from something called "KnowledgeFlow for Procurement." Donald's response on air: "I don't know what you're talking about." Then, after a pause: "So, I've started to put the building blocks together."

What's he most proud of after three years? Not the tech. The daily 09:30 team call. The team is scattered across the country, still here, still moving, working on things that help people who need it. As he puts it: we make a difference.

Two mates and the man who built it. A bar. And all they want to talk about is AI.

Pull up a stool — we'll get the beers in. 🍺

Transcript

SPEAKER_03 0:02

It's the big D That is fabulous. I hope we're recording. We are.

SPEAKER_00 0:12

We are recording. I'm having that. That's fantastic. That is awesome. That is so good.

SPEAKER_03 0:21

Oh, he's gone, he's gone. Come back. Hey Big D you right? Yeah, you? Yeah, good. Thanks. Thank you all. Excellent.

SPEAKER_00 0:32

Well, I didn't get the uh t-shirt memo, but I needed a grey t-shirt, shouldn't I? Uh or something in the middle, something different.

SPEAKER_03 0:40

Grey shows up your man boob sweat most most prolifically.

SPEAKER_02 0:44

I can pretend just just everything to hide everything. Dark colours.

SPEAKER_00 0:49

Let's get this pantomime podcast underway for a special edition of the six-legged pantomime horse with this week with the one, the only, we know him affectionately as the big D. Here we're the man behind all of it. And he is responsible. I mean, look at him, he's cheery, cheery chappy, isn't he?

SPEAKER_02 1:15

Don't call me grumpy for nothing.

SPEAKER_00 1:19

Well, let me let me tell everybody a little, let me tell our audience, I say everybody, our audience, uh, a little story, a little bit of background. Back in June uh 2022, it must was it 2022? Strike it, it must have been. You can't you you said oh um uh fancy a fancy a uh a drink, and I was like, okay, I haven't seen Donald for donkeys, let's go and have a chat. And and you basically said, What are you gonna do? And I was like, I don't know, everything ICT is running along nicely. I think I'm gonna retire. And you said, Don't do that, I've just invented this thing with AI and it's brilliant. I think it's got legs, and I was like, Well, funny enough, Kieran and I are off to uh Vegas for a little jolly in August, and we're gonna check out this whole AI thing, and we went off to Vegas and um got hideously sunburnt, um, managed to avoid um uh any of the uh gaming tables because we saw loads of people throwing loads of money away. Um we avoided a minor incident in the uh Bellaggio uh bar, which will definitely remain just between Kieran and I. And uh and we came back and went, maybe Donald's got a and maybe we should go and have a chat with Donald and let's see what we come up with. And then uh not even a month later, leading AI was conceived and set up as a business, and here we are almost three years later on what has been a roller coaster of a journey, hasn't it? I mean it's been it's it it even though it's three years, it feels like three minutes, doesn't it? Because when you first came to me and said, I've done this thing, and it's like a private chat GPT, we didn't know it was called RAG at the time, did we? It didn't even have a name then. You were so far ahead of the curve that no one else was doing it in the country that we knew about. You were the only one. Nobody wanted it.

SPEAKER_02 3:05

No, nobody knew what it was. I mean, every time I showed it to anybody, they're like, Oh, that's magic, that's that's that's for Eridus. What's that? They didn't believe that they could do what it was doing. Yeah, yeah.

SPEAKER_00 3:15

Cool. So, um, so yeah, Donald, I uh it's tradition for us to uh ask our guest to give a little bit of background, a little bit of history, and one interesting fact about themselves. Didn't give me that bit, did you?

SPEAKER_02 3:28

One interesting fact.

SPEAKER_00 3:28

No, I didn't, I just made that up actually.

SPEAKER_02 3:32

That's the hardest part, one interesting fact. Uh where have I come from? Um, so traditionally I'm a software engineer. I graduated with a software engineering degree, uh uh, which was a long time ago. I came out of having a software engineering degree, doing that for for four years, and then decided I don't want to do that. So I did systems integration and did uh a lot of work on uh Unix boxes, some microsystems, uh 3D modeling for people like Ford. Uh and yeah, and then um and then I got a job doing some integration for Barclays and worked for Barclays for five years or so. Um, so I've been out of the integration piece, but then when you when you're in the integration piece, you do a lot of uh subconscious coding, you do a lot of scripting, you do a lot of uh efficiencies, you do a lot of things, you know. Why would I do that four times when I can just write one script and do it once? So you you sort of you're never too far away from the software engineering part, and then and then, like you said, 2021, 2020-ish, I was sort of tinkering around and thought, oh, what's this new stuff going on over here? I'll have a look at that, and then it just bore out of that, really. It was born out of that. Um, and here we are. I didn't realise it was three years. I keep telling people it's must be about 12-18 months.

SPEAKER_00 5:07

That's because we don't let you out very often.

SPEAKER_02 5:10

Oh no, it's but it's um but the iterate but where we are, I mean we can go into it, but it it's the the the the reason it's sort of seems so short is we're not in any traditional business that's sort of manufacturing that that your product set changes every two, three, four years. This product set evolves and changes every two, three, four weeks.

SPEAKER_00 5:35

Yeah, it's much faster than that.

SPEAKER_03 5:39

Interestingly, Matt Atkinson on the um FPAI side of our business says this is the hardest business he's ever had to be involved with. And is it like you've got customer adoption problems, you know, trying to help them use it. You've got the constant speed with which things are moving all the time. Uh, you've got the kind of customers, I guess, immaturity in AI. I mean, by definition, we kind of all are really, I guess, in that world of where what do they choose to do with AI when so you've got that that makes sales very hard because you you're always up against this kind of, yeah, we know we want it, but we're not sure what for. And uh, and we try and help them with that. So, yeah, it's it's it's interesting. The pace is great fun, I think, but does make it probably harder than most businesses to run, isn't it? We're not just selling the same old thing time and time again.

SPEAKER_02 6:26

It's really difficult, and it it's that's why we we we try and or I try and focus on I try and focus on the outcome, not the technology. Um, you know, the outcome is important for the for the user, for the custom, for the customer, for the company. You know, they have a job to do, they have a you know, and and we need to facilitate that job in a much more faster, efficient way. We don't go in and say we're AI, woo-hoo, you know, how amazing is this? You know, it's about the outcome. But our job is to keep up to date with the changes of in in the AI world and and make sure that we're our tooling is is the best tooling it can be.

SPEAKER_03 7:03

Oh, yeah. Yeah, interesting. And it is, I mean, and that is tough because we're up against all the time, and we we we talk often on the podcast about co-pilot and the claude and chat GPT and I mean Gemini, and we're kind of in a space where we are being squeezed. I think it feels like right now that's squeezing us into a place that only knowledge flow can play, the regulatory space. Maybe that's going to be a good thing. Who knows? It's but it's it's sort of brilliant and awful all at once, isn't it? Is it you think we've carved out a little piece, and then all of a sudden that turns up somewhere else with a free bit of AI overlay on someone else's tool?

SPEAKER_02 7:43

Yeah, well, what we've got is we've got the power of integration, we've got the power to cut across all the systems, we've got the power to mould. You know, we we said we were never going to be bespoke, and we're not fully bespoke, but we are. We have built a platform that's that's adaptable um for every outcome. So we can build integrations into all our systems and pull data from multiple different systems into knowledge flow and then deliver what somebody might take three, four weeks to do in three, four minutes.

SPEAKER_03 8:21

Yeah. Can you remember back to when so uh we started out selling what we now call our version one kit tools? Um, and they so we'd have one for policy, and if you wanted bid writer, we'd have one for bid writer, and in college quality was the first time we ended up in a kind of they needed three in one, so we ended up with sort of clumsily having different links you had to click.

SPEAKER_02 8:42

You then pulled together Knowledge Flow at some point last summer, I think it was about this time last year when Knowledge Flow was first it was out, it was after, or it was even more, even probably during you remember we had the meeting at Tech UK and we had everybody there, and we turned around and said, We we need to move away from a V1, we need a platform play. Um, so we need a platform with multiple assistants that can uh do different jobs, and we looked at it. Oh, I I I described it as you're in a large corporate and you have different departments. So you have HR, you have buying, you have procurement, you have legal, you have all the different departments, and and you need the assistance on the left hand side to be able to, you know, what am I asking today? Am I in, you know, I need a question about HR, I need a question about legal, and I need a question about procurement, and and having that single port to your entire business. Um, and then that's where it came out of. And then you came up with knowledge flow. Um, and then after that, it was just born out of that, really. Yeah. It was like, can I have four assistants? Can I have more than five? Okay, yeah, can you got 10? Can I can I have more than 10? Oh, yeah, you can have 30. Or can I have more than 30? Yeah, you can have a hundred.

SPEAKER_03 9:58

I was gonna say, so one of my favourite things about leading AI as a company has been working with you and getting to know you. Um, and that the your your anything's possible attitude, although not without a fair bit of resistance and moaning along the way, I would like to suggest. But the thing, but in when it in in so the what matters, I think, for our customers, and we were having a conversation earlier with someone about vibe coding, and everyone can vibe code something now with some pretty basic uh tools knocking around. The thing that you've brought from the very beginning, and it was most evident when Ibby joined us in the dev thing, is is that old school, solid, safe, repeatable, documented, tested, uh bringing all of that with it so that we're not we can confidently stand behind the tools. And I think that is really important, but it must be a continual battle because you can vibe code something in 10 seconds and probably spend half an hour having to write the documentation for it afterwards.

SPEAKER_02 10:57

But you've got to you've got to know it's software engineering, is is is it's the foundation of of binary, of bits and bytes of computing. And you know, as much as we all love IBI and we and we do, you know, there's some things that I have a conversation with him, and he turns around to me and says, Well, how do you know that? Well, and I know that because I've got the foundation that I've I've spent 20 years learning the foundation of a computing system. That's never going to change. But what we're doing now is we're layering on top the the fundamental uh foundation of what a computer system does, how it talks to the infrastructure, how it talks to the hardware, and that's still the same. Um, but most people just don't know how to, you know, the value coders of this world don't know how a lot of things work. You know, I took talk talked to Ibby about networking for audio's sake, you know, and and uh I was talking to him about subnets. And he was like, What? I was like, you know what subnet is and how you subnet and super netting and all that. He was like, no. And those are just fundamentals of everything that never changes. Your networking infrastructure, you you your protocols never change or haven't changed for a long time.

SPEAKER_00 12:14

Can I just say has anybody else got volume uh from sound issues? Is it just me?

SPEAKER_03 12:18

I'm hearing I'm hearing so I'm not got sound issues, I am hearing some further interference than as usual.

SPEAKER_00 12:23

There is, there's a lot of interference I'm getting. Right. Is that any better? Yeah, much better. What was that?

SPEAKER_02 12:30

I think it's my fan that's going off.

SPEAKER_00 12:32

Uh yeah, yeah, thanks for that. Um I was just gonna say we we were just uh uh just talking about Ibby. Really interesting little story about him uh because uh he he probably never listens to this, but others others may. But yeah, he um he he was a graduate and um uh couldn't get a job and uh he got like a first class degree, couldn't get a job, and um uh it was actually uh he he paid like a recruitment uh person to to get him a job. And they c they contacted me and said we've got this young person, great credentials, uh uh are you interested? And um and we met him and we thought we'd take a chance at him and now look at him, how he's how he's grown. But it it's interesting, isn't it, that because the the one of the people we we've been talking to on the kind of special guest podcast was earlier was talking about you've got to get the foundations right. If you get the foundations right, the rest of it just follows. Uh but if you don't get the foundations right, the stuff that you end up making is built on sand and dodgy, and how do you know it works, how do you know it's ethical, and all that good stuff.

SPEAKER_02 13:36

Yeah, we spoke about this. How how does you know people who are vibe coding, you know, some of the ideas that they have are brilliant. I never take that away. The the ideas that that are coming out are superb. It's just the tooling that's that's not not great. And the tooling we we we spent such a long time getting 42,001, 27,001, all those are part of our foundation and what we build on. You know, we have to have architecture diagrams, we have to we have to be answerable to our system.

SPEAKER_00 14:11

Interesting enough. I've had a I've had a request from somebody in healthcare who just gone through an ISO uh order, uh a series of ISO audits and go. I was like, if you'd only told me a couple of weeks ago, we could have we could have set that into your knowledge flow for you, and you could have had one of them, and they were like, Oh!

SPEAKER_03 14:26

Yeah, the ISO tool is a very helpful have it sitting in a monthly ISO meeting. I am often looking at the knowledge flow's ISO uh assistant to get to get the answers. It's brilliant though, because it knows all the documents we've got. You know, I mean so actually, most of ISO is have you got a thing that describes this and have you got a log of it? And can you prove it?

SPEAKER_04 14:45

You prove it.

SPEAKER_03 14:45

So being being able to ask and get it instantly rather than I mean, God, with my filing system or or lack of filing system.

SPEAKER_00 14:52

Yeah, but your stuff you keep in Dropbox where you shouldn't be keeping it.

SPEAKER_03 14:56

I don't keep any clients up in Dropbox, it's not so because it's uh it's not secure enough. Dropbox make no make no um uh claim to keep your stuff entirely secure. Well, talk about security, Donald. I I would like to explore with you the conversation about building inside the client as your tenancy, because I think we're still unique in that. There are obviously companies out there that will build you your own AI tools that are over to you to look after. Good luck to you after we've left. That is the only other people I think that build inside your Azure, but we've chosen to do it as a kind of as a service. So tell me about that. Why why that and why not more SaaS based just send the data to us and we'll we'll have it all should be far cheaper, a hell of a lot easier. But why not? Data security, I think.

SPEAKER_02 15:46

I don't know whether it's cheaper or easier. I I we started this with security as as the number one piece. So where does your security lie? Your security lies in your your Azure tenant, uh all the layers are already present there. You are not you're not relying on anybody else to feed your security, so your your security is already in place, already there in Azure. Your documents are already secured under role-based access and everything else that comes with it. So why would you decide to break that security barrier and pull your documents outside of there to make it easier to write a piece of software? Surely it makes more sense to write the piece of software that that that complements your existing security strategy, doesn't break it. So that's that's where I was really, and that that's the foundation of where I was. I I still believe that's the right way to go. I still believe that SaaS is is it has a place, but not when you really do want to put sensitive documentation in there. You know, your sensitive documentation could be anything from board minutes to CVs to medical notes to to anything. And are you comfortable with with holding all that information somewhere else? No. You know, you you already have the checks and balances in place on in your ex as your tenant. Yeah. So it's it's easier for you and more secure for you for us to write software that works with you.

SPEAKER_03 17:27

It does take a bit of persuading sometimes that it's a secure option, doesn't it? The amount of due diligence we have to jump through to get access to people's as your tenancies.

SPEAKER_02 17:35

Oh yeah, you do, but then but then that's just imagine on the flip side of that, if you're asking that those those people who have so much security, oh, and can I have all your data please? That you know that the the amount of hoops you would have to jump through to get the data out of that scenario would be tenfold.

SPEAKER_04 17:56

Yeah.

SPEAKER_02 17:57

So I I'd rather jump through their their security hoops than then come and uh you know, and us to take all the document out of that security um posture.

SPEAKER_03 18:09

And it's definitely been a great USP for us in terms of being able to handle sensitive and secure uh private data. Because it's obviously a load of our tools play in that space. So it's uh very cool. The other thing that it would be great to uh get your reflections on is was the admin console. So we met one of our old dev uh friends a long time ago, Neil and I, on the podcast, Mark Slater, Slash, um, as is as he likes to go by. Um, and he we we were it was a pure show and tell, really, and kind of quite open and honest. And at one point I screen shared with him the admin portal. And he at that point went, Holy, that is amazing. He said, like the the concept that you can deploy a tool from one to another without the data, let's just be really clear for our audience. It takes the settings, and if we've built a time-tabling thing somewhere, we can move the time-tabling thing somewhere else. But that idea that that you've built it everything as a variable that we need to play with. So, yeah, how did that why why was that is that just old school software engineering? Course you do it that way. Is it that's simple for me?

SPEAKER_02 19:13

Old school software engineering, that is never building a platform which isn't once you hard code something, you've you've you've narrowed your audience. So for me, everything that I want to build, everything that I build has the ability to change um change its persona, change its its wrapper. So um yeah, it it's I I think that's probably one of our unsung heroes is the admin, the way the way we we deploy. Because we deploy the same piece of software, we maintain the same piece of software across all our tenants, so that the the the core software is the same, but we have the ability to make it look completely different, act completely different, um, based upon again the outcomes that the customer needs. What does the customer want to achieve? Well, we can make that happen, and also the the there was a part of me at the time that thought, you know, with with people's adoption of AI being like a people are not sure what it does, how it works, what do I prompt, how do I say it, how do I put something in. You know, I think one of the first ones I built was was for a um uh it's for the insurance company, uh, and a large marine marine insurance company, and they turn and one of the prompts that I saw some come through saying somebody asked it, what what what ice cream flavour do I like? And and you're like, so you don't know what you're doing, you don't know what it does. So, how are you gonna ask it the right question if you don't know what it does? So, therefore, we're gonna have to help you on that journey, and the only way I can help you on that journey is to be able to configure it and give you the help that you need by the buttons, by the quick actions, by you know, by the prompts and by you know, helping we helping you prompt your way through to your outcome.

SPEAKER_03 21:16

Yeah. Nice. And you've on top of that built the safeguarding piece, which we've talked about on here before. This kind of this is so one of the if you want to put anything in front of students ever, then the Department for Education regulatory requirements are that you have a safeguarding filter and can track and trigger uh if people are having conversations that you would want someone to know about.

SPEAKER_04 21:40

Yep.

SPEAKER_03 21:41

So you built that, which is great because that runs across all of them again through the admin console. It can it can trigger off anyone anytime.

SPEAKER_02 21:47

Yep. And and the the one of the thoughts that I had was that was you know, you have you have edge filtering devices which catch um responses from the internet which are from particularly spurious websites or URLs or whatever. And I want I wanted to go one step further because I wanted to to be able to catch not catch, catch is the wrong word, but sort of filter out somebody asking the question. What is the question they're asking? Because the answer is bad, but the question is also equally as bad. Now most of them catch the answer coming back. So the answer coming back might be something quite spurious, but I want to know what question have they asked in the LLM for it to be able to answer or try and answer that question. Now sometimes and we we you know if you're trying to ask I don't know what what's what's reasonable to say if we have all sorts working in education but if you just ask how to make a bond you know the LLM will not answer that because it's safeguarding it's guardrailed. Yeah. So the edge device that's filtering the the the the request will not pick that up because the LLM comes back and says I can't answer that question. Yeah interesting. But we know that that question was asked.

SPEAKER_04 23:11

Yeah.

SPEAKER_02 23:12

So therefore we've caught it before the LLM can do anything. Yeah interesting. So that's why it was built that way. You know we could have we could have we could have just relied on the edge device but I don't I didn't think that was enough.

SPEAKER_03 23:26

Yeah. And you said to me at some point about the risk with I think it was Copilot but other tools that allow you to tunnel through your firewall to get to the internet.

SPEAKER_02 23:35

Oh so so that yeah so what we have is we have a a browser extension. So we have a browser extension which sits inside knowledge flow. Now with it with we're not trying to overcomplicate this if you use Azure APIs or if you use open AI APIs it will use Microsoft Azure's internet connection or OpenAI's internet connection to retrieve the answer. So it can effectively circumvent any firewalling or any filtering that you have at the edge of the network. So what we have is we have a browser extension which effectively uses your local machine's browser capability through your firewall through your default gateway through everything what you do so any internet received information data whatever through knowledge flow will come through your firewall so it'll be secure through your firewall so if you do it through Azure Azure uses its own internet connection out of the data centers and it doesn't do interesting so we're even more secure than Azure that's what I'm hearing. Yeah we try we try yeah we think about everything you know this is about the software engineering part you think about all the scenarios you try and think about all the different connotations about how who's going to use it how are they going to use it what you know we security safety is is key to these things because if you're not secure and you're not safe then what are you you you know you're just throwing out a chapter. Yeah definitely no we uh now tell me this you've just been redoing the way we do indexing so the knowledge flow um just to quickly catch up anyone who's listening that might not know all about it um it's mainly a retrieval augmented generation of rag AI tool that means you can have a vast document library and it will find answers in there so it's able to answer a question about policy or a question about uh from all of the legal cases that have ever existed on a subject is able to find best practice in offstead reports etc etc but that process of doing that is a process called embeddings and it is more complicated than it appears isn't it you can just press a button and make it do basic embeddings but then your answers will be not correct but you've just been working on a load more indexing can you share that or is this still very secret source we don't want to give away to anyone else that's in this struggle too transparent but the the the problem with vector databases is what we thought especially when you're looking across policies and documents who have so many similar idiosyncrasies or some paragraphs are so much similar or you know they can be you know you can take minute meet meeting minutes that from one minute from one month to the next they're not that too dissimilar because you for you you know the meetings follow a structure anyhow so it's very easy to sort of look at one particular document and go well that's the answer when it isn't the answer because it was it was a meeting from last month not this month um so when we're doing the indexing now we we do a uh we do some pre-processing so we pre-process it so we we're pulling out key vector or key key pieces of data that we will embed into the vector database into the chunks that will allow us to categorically turn around and go ah yes that is the document we need and that is the document you need to retrieve the answer from um so yeah it's it's it's we're expanding on that uh uh quite a lot at the moment you know and all the other bits hybrid search that you have on there NER BM25 re ranking cognitive sentence matching I mean it's like I'm amazed we get an answer back within an hour from them.

SPEAKER_03 27:36

I don't know what some of those things are what are the some of those it's some of it's secret source Neil you can't everyone to know that the combination of those and the settings that you apply to them is what makes knowledge so accurate.

SPEAKER_00 27:48

I'd use that like the prep sandwich where you they'll just give you the list of ingredients and then yeah and you make it and you make it wrong. You can't deconstruct it here no very uh so what would you say is the hardest uh what what two I have two questions and they're very simple but they're simple to ask I'm not sure if they're easy to answer.

SPEAKER_02 28:09

But what what's the hardest technical challenge you've come across in building knowledge flow and what's uh and is that the uh what you're most proud of in what you've done so far I thought it was simple to ask yeah it's it's a it's a tricky one tabling the hardest thing the hardest thing has always been my my nemesis is um uh structured data so excel spreadsheets CSVs SQL the reason it's the difficulty is because what we tried to do is when we when we started this and your your chat GPT your clauds all the all the grocks and everything you know when you when you log in and you try and upload anything more than 20 documents it goes no we can't do that you need to upload you 20 is our maximum or whatever it is so we turned around and went right we can't do that because when we're looking at medical notes when we're looking at solicitors notes we're looking at all the bits and pieces you know those those those case packs can be six seven hundred documents effectively that's easier than reading 5000 lines of a SQL database because 5000 lines of a SQL database the commonalities between those lines are are are so minuscule so pulling out the differentiators between line one and line 427 or whatever that may be that's difficult and that's tricky and there's always a scenario and what I find is there's always a scenario where you feel like you've got to the to the magic answer and you know your your ones all waving in you done it works and then somebody will come along and go just a minute I've just done that and that's the answer and that's the and you just think well I know why that's come out as that answer. I know what it's done but that's just another scenario that I need to cater for to make sure that it works. So this unstructured data is actually easier than structured data in a rag because rag or LLMs are almost better built for unstructured data. Yeah and this is the problem of as we've talked about before is the the the problem for an LLM is it it it knows the difference between a keto and a car but ask it the difference between number 10 and number 11 and they're basically the same thing is then that's where it struggles with numbers entirely doesn't it and you've managed to you've managed to hack that is that the right phrase probably not hack is probably not a word we should use engineer engine engineering but what's what are you most proud of that's what Neil asked the second I think um definitely I I'm gonna be sound sound sad and and and but we're we're the company it's the most proud of actually you know we we all work together we have a call and you know and you two have got to thank for this because it's something I was alien to but we have a call at half past nine every morning with everybody and we all work separately we all work in different parts of the country um we we rarely get together but we should do it more but it's just you know life gets in the way and work gets in the way and we're still here three years later and we have a lot of customers on our books and we are moving forward at an incredibly fast rate and it's it's so enjoyable. You know it's it's grateful. You know to to work in some ways grateful is a blessing.

SPEAKER_03 31:58

That it's good and I love the idea that what you know nearly all of our customers are doing it are focused on helping people helping underserved people in the main and disadvantaged folk and that always gets me out of bed.

SPEAKER_02 32:12

So working with a great team on stuff that matters not just we're not just there to make someone some more money which makes quite seems we do make a difference and it it it's it's making a difference to people's work life at least but that work life makes people makes a difference to people on the ground.

SPEAKER_03 32:30

So here's the challenging questions not this was not technically challenging strategic we had a strategy session on Monday which was really good and we could talk about some of the things that came out of that if there's time but um so I had a feedback session with a customer who has had BidWriter and they had it on a trial basis and headlines are they've come back and basically said we it's not really doing it for us. With we think we're going to hand it back and rather than well they are going to hand it back we'll not think and um it was a difficult thing to listen to but there was good feedback that they were giving there was a whole bunch of it's personal feeling rather than anything which is very difficult as you know a lot of we get from people as well I didn't like the way it wrote like that and we can change the tone a bit but then you'll get another colleague over here that says I like the way it writes so that's always a battle. But here was the thing that went a couple of bits that went through my mind is it strikes me we were we created BidWriter more than two years ago probably when no other tool was able to do anything like that and people weren't even aware that that was something AI could help you with in any you know people were using ChatGPT to give answers to to bid questions but not bringing their whole library of case studies that they've organizations created and roles and methodology statements and all of that stuff and drawing on that. So the difficult question is are there tools we should think about retiring because what my concern is that these guys were using Claude and basically saying you know we're using Claude alongside it and frankly it's a bit better at answering than knowledge flow is and I was like well that's partly annoyingly because Claude writes better and it'd be great once we once Microsoft finally bloody allow us to use Claude inside Azure securely then that could help but the other side of it is that they aren't that bothered about the security of their documents because it's these are pre you know it's a small company we're talking about they've got a lot of proposals the only part of those proposals which is in any way sensitive will be the cost tables and ultimately you know if they if they give all of those to Claude and they end up in an LM that's not a declarable data loss. That's just a I might get told off by my boss. So the question on my mind is given knowledge flow's focus on regulatory stuff where it matters where the answer has to be right and where you need consistency should we think about retiring BidWriter and maybe some of the other early tools which people are now able to do in other ways. And my fear is always that if we keep going head to head with it and we can keep improving BidWriter and for some organisations it's perfect.

SPEAKER_00 35:15

I don't want to lose it so just to be clear I've just I've I've I've done a couple of bids recently and it to get to get 60% of the we're there in two hours where it would normally take you two weeks just gives you such an advantage and it's not perfect. And I think there's something about expectation management for people where but the honest truth I'd be gutted if you if you shut it down because the next time I need to write a bit I'm gonna be bloody starting again and and messing and and actually and the other thing I would challenge back to you Kieran is uh there's something about um I I and and I can and I can and I can cut both sides of the coin right I can be ruthless and say that's that's no good let's just cut that we're not gonna do that that's fine. But on the other side it's driven a whole load of innovation. So for example we did a bid for an organisation and um and they asked they they basically had two spreadsheets with a hundred and ninety questions in and what a pain in the bum that was and I was cutting and pasting and cutting and pacing and I was sticking the prompt in asking getting to get an answer sticking it in the bloody spreadsheet and then Donald went oh I'm gonna I can fix that so he'd let me spend two days doing all of that kind of cutting and they oh yeah I'm pretty sure I can fix that the next thing he comes on yeah press this button and it do do do do do do do do yeah answer we've now answered question seven of 197 we've now answered question eight and and you can watch it just go through and just to get that it is is incredible. So I hear you but I would I would advocate the bid writer isn't one of the ones for the cult or if it does get cult for up for other people then you know I think it's a personal choice. I think if anyone's using it and they're getting value from it then brilliant so I I mean I'm I'd begut it but it drives innovation and that that that that's the point that I'm trying to make is that it's forced it's forced us to do things and and not and not only do did did we now move from pressing the button Donald's created a little wizard at the front which says you know how many turns do you want it to do so you'll get the answer back and then and then it can ask it can ask itself the question you know is this a good answer what score would you give it if it's less than four out of five then go back and have another bash at it. So just that not accepting the first answer out of the door um uh and actually providing some judgment on it and so it's already moved on way way much further and I think for for some people I get it and we we've talked to lots of people about this recently that whole piece about um uh people using the frontier models that are are free and accessible they're getting or or you're paying you know twenty uh twenty dollars a month or whatever and you're getting access to the latest bells and whistles of course they want that but the reality is uh you can't have them in the secure environment in the UK at the moment you just can't and um so it's managing uh some of it is managing expectations as well and and that is a real challenge for us it is a real challenge and I don't think that's gonna change I don't think it's gonna change anytime soon is the honest truth. I think we're just gonna have to keep on going but to answer to answer your question should we be uh retiring some of our products quite possibly but we we're already doing that we're doing that with the version ones that you mentioned earlier you know they're going out we're changing we're updating the models but this another thing that we do we just don't tell our customers we we move models you know they're gonna they they most of them don't even know which model they're on is the honest truth so all they want to know is it actually works it's a bit like hygiene in it they just want to know that it it's working for them so anyway I would I'm gonna stop advocating for bid writer and let Donald answer the question that you asked him 10 minutes ago so my my my simple answer is what we started off as saying is uh the speed of change and the speed of improvement that we are experiencing in the AI world makes bid writer always going to improve so yes you are up against somebody who is capable of using Claude or Claude code or codex or whatever they're using and they're using it in a in a in a way that is that is that that works for them great um our bid writer will always improve so the different the newer models the newer prompts the newer sort of you know tooling that we put put around it that will always improve and I'm really rubbish and crap at metaphors but you give me a trowel and a and a and a pile of bricks and ask me to build a house you imagine what it's going to look like look like so basically there are people out there who can prompt and there are a lot of people who can't so bid writer cuts across the masses there are a few people out there who do know how to prompt and then do how to use it but the majority don't yeah interesting and I think I mean one of the things on my mind in a slightly more technical answer to BidWriter's problem is we've designed Bidwriter pretty much well really well for a kind of public sector procurement audience where they want lots of evidence of how have you done this before.

SPEAKER_03 40:41

And so therefore if you think about and in a technical sense what we do and built by design is that as soon as you're answering question one, it's going to do a retrieval into your index bring back a load of chunks that are relevant stuff and write about those in its answer. And I think one of the things from this feedback from uh this company earlier is they want a bit more time on the kind of creative answer of course because that model will always favor uh going back to what you've done already as opposed to some new way you could do it. So I do think there's and this is the stuff of a button prompt so it's pretty straightforward to test is to say to it first up let's do some brainstorming or you go away knowledge flow and give me some thoughts about how I could best answer this and then let's go look in the index for stuff that's adjacent the same as or whatever.

SPEAKER_02 41:33

The personality of the the LLMs as well you know we know the personalities of the LLMs can change change the answers or change the inference or change the way it writes dramatically yeah so it's for us to know the personalities of the LLMs to put them in the right places for the right people.

SPEAKER_03 41:49

Yeah.

SPEAKER_02 41:50

And we can change you know going back to the admin console if they turn around and say we don't like that with a press of a button we can change all of their their their um knowledge flow to five five five six sol five four four zero whatever you want to do it five six yeah in as yourself securely can we uh I don't know I'm not checked I'm not checked yet I checked mod route yesterday but but I've not checked five six I bet it's not there yet we'll probably get it in about next September 2028 I imagine they'll release it to the UK south they're rolling them out really fast now because it's a it's a it's a race to the to the top isn't it we'll race to the bottom once they let us have Claude then we'll be laughing well hopefully I don't know why that I don't know why there's not a the the there's a blocker on that one it's there it's listed but you just can't get access to it.

SPEAKER_00 42:39

Yeah maybe uh maybe Microsoft haven't heard their talk and Billy that's what it is.

SPEAKER_03 42:44

Yeah. So on the Monday Donald this is a this is a good share with our audience about how Donald works. On Monday we spent the afternoon together talking about where we should go the importance of regulatory uh the regulatory focus for knowledge flow evidence packs audit trails more data integration layer all that good stuff we also talked about building a knowledge flow for procurement that could do a bunch of things and inquiry handling and supplier sorting and invite quotes and deal with quotes and effectively automate the whole kind of first steps of a procurement world. So you you said okay great I haven't done anything with that yet. Well so I get alerts as you probably know although you're probably going to turn them off me. I think you Ibby and I are the only people that get automatic alerts if one of our tools goes down is offline or whatever else. And I must say I've only ever got them from sandbox uh tools I've never seen one from a customer which is great that they're always operating. Anyway I started getting one from a thing called Knowledge Flow for procurement I've I've had a few head ups for this thing that's going on in offline as you're presumably doing something. So do you want to share anything about that Donald I don't I don't know what you're talking about. No idea.

SPEAKER_04 44:02

No idea.

SPEAKER_02 44:03

No, it yes, all right. So I I've I've I've software engineering starting, I've started to put together the building blocks. There you go. So that's that's all I will say is that this isn't this this is this is not this is building blocks detailing, you know, the detail spec I've got so far, it for me is probably four or five pages long. About right, I need to go for this, but that needs to be able to do that, that, that, that, that, that, and then then you build it's building blocks, it's like Lego, that's all it is, and you can't say Lego, and other other building blocks are available. But but it's that it's building blocks. So I've started to put the building blocks together because again, we know this. I'm I'm a I'm I'm an engineer, I'm not a designer, so I I um I can make things functional, but you ask me to choose a colour, no idea. Black, black. That's all I thought I just want to pick it up. White and black and white, yeah, yeah. Maybe that's what we get on Donald's.

SPEAKER_00 45:13

Black and white. There's been plenty of times when you've gone, oh, I think I've really pissed him off because I've asked him to do these 800 things, and then Donald will go, no, no, you can't do that. No, no. Yeah, yeah. And the next morning be like, here it is. Here it is, so you do have prior Donald. I didn't sleep last night. Well, there you go. That's right, I didn't sleep, but here's your two or four, yeah.

SPEAKER_03 45:36

There you go, we've fixed it. And timetabling is one of those which um you've and I've tried to explain it briefly, and hopefully not saying too much that is uh secret, but what you've managed to do is blend LLM and effectively old school algorithms. Yes. And that's how you're making timetabling work.

SPEAKER_02 45:53

Yeah, there's there's there's there's a pro the the process, like again, the building blocks. There's one building block that might turn around and say, I need to this needs to be hard-coded because I know what the inputs are, so the inputs need handling in a particular way. But then some then then you need to move it on to a stage, and the stage two is well, actually, somebody might mean that, or somebody might mean that, or somebody might, you know, somebody might call a course this, or somebody might call a course that. So, you know, to to to to code that it's not impossible, but your your statements or your functions become unwieldy because you're trying to cater for every single possibility, even spelling mistakes. So using an LLM in that scenario, weaving the two together makes the whole process incredibly functional and incredibly powerful.

SPEAKER_03 46:44

Yeah, it is very good, and it's doing a great job of so far. And we're looking forward to getting into the wild to get some people actually using it. But it I've shown it to three colleges and a university, and just you know, the the initial feedback is like the old days of showing them rag for the first time. You get this kind of that's amazing. When can I have it? So that would be really good to uh to get out there.

SPEAKER_00 47:08

Um, yeah, good. Cool, cool. Where do you see things going from here, Donald? What's next? Where do you think what do you think the evolution's gonna be? Where do you think we're gonna be in another three years' time? Three years is a long way away. I mean, we couldn't have predicted this stuff. We we couldn't have we couldn't have predicted. Three days time where we're gonna be 40 by, huh?

SPEAKER_01 47:37

I don't know. I think I'm not gonna change my mantra. My mantra is it's about the outcomes, it's about the customer outcomes.

SPEAKER_02 47:49

So the exciting part for us is you don't you wake up in the morning and something else happens, something else has changed, something else has been released, some of the functionalities here, you know, and we're not building the LLMs, we're not we're not we're not in that game. We are we're in the game of tooling, we're in the game of making you know pulling all the tools together and making all those tools work for our customers. That's the game we're in, and and it's exciting because there's there's all sorts of tools out there. Um so where we're gonna be in three years, I hope we're still here. I still hope we've got a hundred times more customers, and um yeah, and we're still doing what we do because I enjoy it. Uh, and and that's partly, you know, we we're all old enough and ugly enough now that we need to enjoy what we're doing. Yeah, we are.

SPEAKER_03 48:37

Yeah, I think we have. As we said, I've said a few times along the way, is the good thing about sort of being slightly uh slightly more mature in your career is that you basically can rule out, we don't work with arseholes. I try to I try to sift out customers who are going to be arseholes and just don't really carry on with them or make it too expensive to make them go away. And it's quite it's quite nice to be able to feel that you're not like I remember in my sort of like late 20s, early 30s when you just had to just gunner everything, and it was always oh god, this is really a difficult customer. Now you just think no, they're just an arsehole. That's all they are, and don't work with them. It's a fulfilling ethos getting rid of our souls. Well, there you go. Maybe there's your uh there's your uh podcast title, Neil.

SPEAKER_00 49:24

Getting rid of our souls, getting rid of our souls, brilliant. Uh, gentlemen, it's been emotional. Donald, thanks for making the time. Pleasure, thank you. It's it's been great as always, and uh, we'll get this one out. And hopefully, our audience will have learned something really interesting along the way. Uh, and if you could tell us what it was, then that'd be brilliant. So thanks very much. Have a great day.

unknown 49:45

Bye.

SPEAKER_03 49:46

Thank you everybody.

SPEAKER_00 49:47

See you bye.