From Skeptic to Superpower: Real‑World AI Coding Workflows That Scale | BRK229

0:00Priyanka Sharma: Hello, everybody.

0:03It's so nice to see you all bright and early.

0:06Thank you for coming.

0:07I'm Priyanka, and this is Mario.

0:10And we work at a company called Thiink,

0:12which is in the energy management space.

0:15In the past, I ran the Cloud Native Computing Foundation,

0:18CNCF, the home of Kubernetes.

0:21That was the last cool thing before this cool thing.

0:24And in the 10 years I was involved with Cloud Native,

0:28I saw the same twists and turns with enterprise adoption

0:32in the Kubernetes landscape as we are already starting to see

0:36in this new world of generative AI.

0:40For the past six-plus months,

0:41I have been vibe coding using generative AI to build.

0:45And I'm really excited to share my perspective

0:48of the experience, the similarities with the past,

0:52and where the deja vu comes from with you.

0:56Mario Toffia: So, my name is Mario.

0:58I come from Telco and IoT background.

1:032022, I started coding Energy OS, it was called,

1:10where we created a platform that was enabled

1:15for LLMs to execute on.

1:17So back then, I saw that.

1:20And the epiphany was 2021 when I tried out GitHub Copilot.

1:25I don't know if you remember that back

1:27in those days, tab completion.

1:31And I realized within five to 10 years,

1:3580% of the simple code will be done by AI.

1:41That's a given.

1:43And we're here now.

1:46We surely can do that.

1:47We didn't think to do the simplest tasks.

1:50That's 100%, I would say.

1:53Priyanka Sharma: Right.

1:55So at Thiink, we power over 10,000 buildings,

1:59both commercial and residential, in the various needs of power.

2:03And so, some parts

2:04of our technology are critical infrastructure.

2:09Mario Toffia: Yeah.

2:11I should be coding now since -- yeah, yeah.

2:15Priyanka Sharma: That's true.

2:16Well, okay, Mario, do you think you'll feel better

2:20if you just set off a process

2:21and something starts getting built?

2:23Mario Toffia: Yeah.

2:23Priyanka Sharma: Okay.

2:24All right.

2:24I just took the liberty of pulling

2:26up your to-do list for you.

2:28And so, if you want to go ahead

2:31and set something off, feel free.

2:34Mario Toffia: Yeah, absolutely.

2:35So, what I have to explain first is that, so you get accustomed

2:43to it, I will go through this during the presentation.

2:46But I need to kick it off because it will take some time.

2:50It's an autonomous flow that do gating

2:53and fork out lots of agents.

2:54But what we do, basically, we have PRDs that break

2:58down the features, the features break down into tasks.

3:01And those tasks we can implement in waves in parallel,

3:05depending on the dependency graph.

3:07And we're going to kick off one of the implementations

3:12that hopefully creates freeways.

3:14I don't know because it's an LLM.

3:16It might even shop it another way.

3:19But hopefully it does that.

3:21Priyanka Sharma: All right.

3:22Mario Toffia: So, can you see my screen?

3:28Yeah. So, I kick up Copilot.

3:34We're using the Copilot CLI.

3:36I run it in autopilot so it doesn't get stopped.

3:46So, let's see.

3:53So, what we have in our repo, we have all this information,

3:57so we can we have a custom command that we can list all

4:00and the status of them.

4:02So let's, I have a sheet one there.

4:08So, we use a skill called implement feature.

4:10That's the orchestrator that forks out

4:13and uses this excellent task scheduler in this Copilot.

4:18So, you can actually schedule sub-agents

4:20that can schedule other sub-agents

4:23to handle the context really, really well.

4:26So, let's fire it up hopefully, because we have got lots of time

4:31out since Build started.

4:34It will finish up.

4:36I have a backup if it doesn't.

4:40But what it will do now, it will read up the context.

4:44It's immutable, everything, so it will read back,

4:47see which tasks belongs to this feature,

4:50what the peer D it belongs to, and read up the architecture,

4:56and then start to create a plan for the implementation.

5:00And switch over.

5:03Priyanka Sharma: All right.

5:04So, do you feel a little better now?

5:07Mario Toffia: Yeah.

5:09Priyanka Sharma: Awesome, awesome.

5:11We weren't always this comfortable

5:13with AI-enabled workflows.

5:16It all started in 2025, really, in earnest, in Thiink,

5:20when our CEO, just like I think many other CEOs out there,

5:24made the mandate that we should all try to use AI-enabled tools

5:29to help us do our work bigger, better, faster.

5:32Unlike most other CEOs, though,

5:34our CEO was utilizing these tools himself

5:37and had a little bit more credibility than normal

5:40about what was working, what was not working,

5:42and he truly saw the potential.

5:44At the same time, we as the team,

5:45though, were not so stoked.

5:49As I said before, 10,000-plus buildings are relying on Thiink

5:53for critical infrastructure, such as heating, ventilation,

5:56and orchestration for that.

5:59You don't really want to be the cause that a pensioner

6:01in northern Sweden didn't get heat

6:04and then had health problems or worse.

6:08Mario Toffia: It can be really, really cold.

6:10We have, where I come from, near the polar circle,

6:13we have minus 40 degrees in the winter, Celsius.

6:16Priyanka Sharma: Yes.

6:17And we're deployed in places like that.

6:19So, that tells you something.

6:22But we gave it a go with caution.

6:24At first, our efforts were more POCs that were glitchy.

6:29I personally felt like I was wrestling hard

6:32with the early cloud models of Sonnet and then Opus

6:37to build a new website for Thiink.

6:39It was tough work in the beginning.

6:42I felt we were fighting every day.

6:45Fast forward to February 2026, and I had this major wow moment.

6:52For me, it was with the release of Claude Opus 4.6.

6:56I remember actually messaging our CEO, being like,

6:59I just built this section in five minutes.

7:01This is crazy.

7:02And it was so exciting.

7:04I even remember thinking that I would think

7:07of this moment in the future.

7:10I know for Mario, he has a different experience with 4.6.

7:14He has his own favorite model story he will tell you

7:17in a minute.

7:18But we got really started and started seeing some good wins.

7:22And all areas of business were impacted.

7:25So, our CEO and controller in the business,

7:28they created a Claude-based request for quotes workflow,

7:32which is basically, it's like an RFP process to apply

7:36for commercial buildings, installation,

7:39and electrical projects, etc.

7:41And instead of days, it was taking minutes.

7:44Now, keep in mind, each application is worth

7:47between $100,000 and $800,000.

7:49So, this is really meaningful.

7:54Even in the teams, like our head of product experience,

7:57she has now been prototyping and building features from zero

8:02to really working options in days.

8:07For example, meet Sparky, who is a chat agent that she has built

8:11to talk about what's going on.

8:12And she has a technical background.

8:14She is an engineer by training.

8:15But in the past, having leaned more into product,

8:18she would have felt a little reticent to take this on.

8:21But here, with her guidance, Codex did a lot of work.

8:24And here we are.

8:26I myself have become the one-person martech team

8:30for Thiink, where I'm building the company website in-house

8:33with all the CMS, CRM attached, all of that.

8:38And Claude is completely unbiased

8:40and never flattering me, (inaudible) I'm saving us

8:43about between $120,000 and $200,000 a year.

8:48But this slide, it behooves me to point out, is also a bit

8:52of a double-edged sword.

8:54You see here, I'm showing you screenshots

8:56and not the actual site.

8:58So in our business, we don't use this particular site

9:02for lead gen and those kinds of business needs.

9:04So, I had a bit of leeway to make

9:07this as beautiful as I wanted.

9:10And so, here we are, six iterations later,

9:13and I really need to ship it.

9:15So, generative AI is enabling my OCD a little too much.

9:21But on the other hand, you see those case studies and news.

9:24This would have taken me weeks to build

9:27with all the permissions and stuff.

9:29And this was two days with final approvals.

9:31So, it's pretty cool.

9:34Mario says that on the engineering side,

9:36the 10x return comes where the engineers are weakest.

9:41Like someone who is not security-minded will suddenly

9:43have better security posture in their PRs.

9:46He himself is a perfect case in point.

9:49If I'm not wrong, Mario identifies

9:51as a back-end engineer.

9:52And as you see in his GitHub repo,

9:57he says, here's an ugly mock.

9:59Now, sure, this is not a Picasso,

10:01but this is fully functional and going to do the job.

10:05This is a great UI to start from,

10:07and this is generative AI enabling Mario.

10:10So clearly, Thiink was seeing lots of wins

10:12in different parts of the business.

10:15We were really having what can be described as a good time.

10:19What could possibly go wrong?

10:25Well, so this is where we come to new tech, old problems.

10:33On April 19th, 2026, Anthropic banned our organization.

10:40All our AI-enabled processes came to a halt

10:44that were related to Claude anyway.

10:47We made a business decision that all the business processes,

10:51such as the RFQ work, even my front-end stuff, we were going

10:54to wait to figure out until things got sorted.

10:58And the core engineering would continue

11:00through another backup tooling, such as Cursor.

11:04Because then we had to make that decision,

11:06because in the other tool, there is a markup

11:08on tokens, and that costs money.

11:10Being a bootstrap startup, you have to evaluate your choices.

11:15So, it was a dark time, which was

11:16very demoralizing for this team

11:18that had suddenly become this group of builders.

11:22And I'm happy to say, 10 days later, we resolved things.

11:26We're back in business.

11:28But even as we did that, we've had glitches to solve

11:30on the billing, payments, all that side, reminding me

11:34of what is the famous thing

11:36that we've been dealing with in cloud native?

11:38FinOps. So really, this has been a very deja vu experience

11:44for me.

11:45What are we dealing with here?

11:46Vendor lock-in, lack of data

11:48and workflow portability, a FinOps nightmare.

11:52This is, I mean, we just got into this new era,

11:54and we're talking about all the things from the old era.

11:57So this, in some way, actually also gives comfort because,

12:01okay, we've done this before.

12:03We figured it out.

12:04We'll figure it out again.

12:08The big learnings we had as a team were:

12:11If you're working in-app, like if you're in the Claude app

12:13or ChatGPT app for something and have built a process there,

12:17there is no real portability available in those.

12:20And things can be truly lost.

12:23You'll get your data back one day,

12:25but you can really be stuck.

12:28Anything that relies on implicit memory

12:30or a specific chat window.

12:32So, for example, I built an agent to help me

12:35with the copywriting for the website and had all these,

12:38like tone of voice, product marketing, and this and that.

12:42And with the one-million-context window,

12:43I was really having a good conversation

12:45about which direction we want to take.

12:47And then suddenly it's gone.

12:49So, I had a lot of friction just internally moving

12:53to a different tool.

12:58Any tool realistically that adds a markup is not a good backup,

13:02because if you have any financial responsibility

13:04in the company, you'll end up having to make choices

13:08when you've gone dark with one vendor.

13:11And not everything can go

13:12because the markups are pretty high.

13:15And this is the most important one.

13:18Brilliant colleagues, whether they're human or machine,

13:22can quit at any time, and you have to prepare accordingly.

13:26One of the reasons generative AI really took off in Thiink is

13:29because we had so much fun building with it.

13:32Here was this really smart person

13:34that was always available, ready to talk about my ideas and go.

13:38And then suddenly they've gone dark.

13:41This has happened before in human scenarios.

13:43It can happen in AI scenarios too.

13:46So, that's my story.

13:47Mario, do you want to tell yours?

13:50Mario Toffia: Yeah, let's do that.

13:51Priyanka Sharma: I'll switch over.

13:53Mario Toffia: I will quickly show you

13:54that it is still implementing, has some problems with the TTI.

13:59Let's see if it solves it.

14:00But it's done the smoke test and so on.

14:05Yeah, so the story is all the software, the drift.

14:12We all have it because we have an internalized view

14:17of how a function behaves or what a word means, and so on.

14:23And we spend a lot of time syncing this in architectures,

14:27design, meetings, and code reviews.

14:30So, we review each other to sync this.

14:34And that works pretty well, lots of job to do.

14:38But LLM is not any different than a human.

14:43It has chewed a lot of data on the internet and got some form

14:46of internalized view of everything.

14:50And we need to sync that to it as well.

14:55But it has a bit more limited context and attention span.

15:01But the thing that it does most is that it can chew so much data

15:07and spit out so much code in a so short period of time.

15:12And it has an excellent way of writing self-confident code

15:20that are not correct by your standards.

15:25And it can even invent stuff from thin air.

15:30So, the drift is real, and we need to harness it in some way.

15:37Because the consequence is that the code base will rot.

15:41You will starting to have slight differences.

15:44You have different words for the same thing.

15:49And then when we start to reason about it,

15:52or the LLM in our case, because we're auto-coding everything,

15:57it starts to interpolate the code wrong

16:00and do even more wrong.

16:02And since it's so fast, it will go really fast in that way.

16:08The bug frequency goes out of the roof.

16:12I saw an investigation of teams that uses AI for coding for 50%

16:21or more, the bug frequency goes up 53%.

16:26And one big thing is that they haven't harnessed it,

16:30so it tries to create the correct code.

16:36Yeah, so none of this.

16:37And it becomes very expensive to maintain.

16:41The thing that should accelerate us will decelerate us.

16:46So vibe coding, I think, is really good

16:48because you can get your prototypes

16:50and everything ready in time.

16:52But if you're going to have a product that's going

16:54to ship next year and the next year after that,

16:57then the problem will start piling up.

17:00So, enter the Thiink Harness.

17:03So, yeah, we have to do as we always have done.

17:06Do the spec, do the design, do the architecture, and make sure

17:10to communicate that properly to the context.

17:13Because it's all about context management there.

17:18And what we started with was we used ChatGPT

17:24to do our research back in the days and iterate our designs.

17:31They have really small context window.

17:34So, we tried to tell it what to do, and it derailed directly.

17:40And it was hard to get any usable code.

17:44Test code, yeah, could do.

17:47Documentation, okay.

17:49But when Claude 4.1 came, yeah, now we can start doing this,

17:56audit the test, but small fixes, and then documentation.

18:02But we realized we have a lot of products.

18:04We have different types of languages

18:07and somewhat different architecture.

18:10Quite sound, of course.

18:13When it's a million-plus lines of code,

18:15you have better and worse parts.

18:18But most of it is quite good.

18:20But we discovered that we need to have a way

18:33to express our architecture

18:35that don't make the LLM derail so easily.

18:39So we, I mean, we have used some parts of domain-driven design.

18:43We have done some parts of CleanArc

18:46and hexagonal and so on.

18:48But now we restated our purpose and restated our architectural

18:53and design principles so we can make our system much easier

18:58to interpret.

18:59So, we even did the scaffolding so it will be easy

19:03for the AI to understand.

19:04Oh, I go to ports to examine an interface.

19:08I go to adapters to see which are the implementation

19:12and the domain context, and so on.

19:14So, that was our first discovery.

19:18The second one was Claude 4.5

19:22because we didn't like 4.6 so good.

19:25I don't know if you remember this, but it stopped adhering

19:28to instructions that, well, things that worked really well

19:32in 4.5 started, ah, I do whatever I like.

19:36And then the override the system prompts for that.

19:39So 4.7, thankfully, was the patch for that one.

19:44But 4.5, now it really could code.

19:48We can have two different services that we are going

19:51to join and do some data.

19:53Previously, that would be crap.

19:56But now it actually could code

19:57and do understand those two services.

20:00Sorry, I'm going to cough.

20:05So, we wrote a lot of skills, a lot of specialized agents,

20:09so security auditor, resilience auditor, different experts,

20:13API experts, and so on.

20:16Earlier, the context we had was too small

20:19to really explain what we want to do.

20:22Now our context was too large because now it's trying

20:27to pack everything into the context.

20:30And what should it pay attention?

20:32Because LLM is an attention mechanism in the end.

20:37That's the transformer architecture.

20:42And then we realized that we need to pack in information

20:46that it should be guided by,

20:47and then what should it pay really much attention to.

20:52So, we don't have must in everything and must not.

20:55We have a few of those.

20:56So, it really paid attention to that.

21:00And a continuation, then we started

21:06to continue develop this.

21:07And we will develop this in the future for as long

21:11as we develop our software.

21:15But it's a better harness.

21:19So, we started to think that, okay, let's put our information

21:25into the repository in a more machine-interpretable way.

21:31You can RAG, of course, but RAG is more, oh, I can take this

21:35and this and this because you use some vector similarity

21:38to find the text.

21:39And you can do clever things.

21:42So, you can use a graph database with RAG.

21:45And there's lots of job with that as well.

21:48So, let's stick to good old strict traditions

21:53to define our specs and make it just adjusted a bit

21:58so the LLMs can easily show it and put it into their contexts.

22:04We have a product owner in our company that is awesome.

22:09He knows everything about our business.

22:12He knows the customers really well.

22:15He has constant contact with them.

22:18And he knows technology okay.

22:22And what he does, he is chaos engineering.

22:27So, he is really creative.

22:29So, lots of Google Docs.

22:31We have a box that he does with Claude to try

22:35to visualize concepts that he wants to have implemented

22:38and put those into roadmap items with slapping

22:41into a few goals and some use cases.

22:45And then we have to break this down.

22:47And what we do now is that we use a skill

22:49to break those down into PRDs.

22:51So, we define those scopes.

22:53And then we take those PRDs, take one of that,

22:57and use our tech skill.

22:59And that tech skill, it automatically breaks those

23:03down into that skill features that we can parallelize.

23:10Sorry, I'm a bit nervous.

23:12Tech feature, it's called parallelize.

23:17And it uses several sub-agents depending

23:21on what the PRD is about, what we're supposed to do

23:24in this feature, and writes one first.

23:29But it is a human in the workflow.

23:33So I, as an engineer, steer it and say, okay,

23:38but you got this wrong.

23:39You should partition this way.

23:42We need to do this.

23:43Have you thought about that?

23:45And it reiterates from the older iteration, the instructions,

23:50and see, okay, I need help with security and resilience

23:54because I didn't cover that good enough or figure it out.

23:59And then when it's done, we have a feature draft.

24:02After the feature draft, then the --

24:07Priyanka Sharma: Socratic method.

24:08Mario Toffia: -- Socratic method, yeah.

24:10So, Matt Peacock's excellent grill me.

24:14If you haven't tried it out, try it.

24:17It's awesome.

24:18It's a few lines of code.

24:20But it's a Socratic method.

24:21And so, heavily inspired by that,

24:25we use our context, read it in,

24:29grill the user on our common terms,

24:33ubiquitous in the DDD world.

24:37Is there any architectural changes that we need to do?

24:41And so on.

24:42And I have to answer every question.

24:45It forks out sub-agents

24:47to perform more grilling in specific areas.

24:53And we were keen on security, so often it forks

24:56out the security agent for that.

25:00When we're done, we have written a brief.

25:02We have canonicalized keywords, and we have ADRs.

25:11And then the last step is to rectify it.

25:15And it's when the last QA sync,

25:18where we make sure that we are all set.

25:22So, this is the QA gate, you can say.

25:25And then it will write the ADRs, write the (inaudible).

25:28Because from now on,

25:30since everything is immutable, we can't change it.

25:36And that's why we do that.

25:37It's because the LLM should be sure of I can read this path,

25:42and I don't need to care anymore

25:44because it's immutable, all this data.

25:48So, what's executing right now is the yellow,

25:52and we split those into tasks.

25:56We split it on the DDD boundaries,

25:59hexagonal boundaries.

26:01So, we have a step for a human to say, okay,

26:06these tasks looks good, had never adjusted them.

26:08So, this will be the next flow

26:10that we skip just automated totally.

26:13Split up all these tasks.

26:15It's a bag, so it's topology sorted so we can make sure

26:21that dependencies are executed in the right way.

26:24And we can parallelize it to make it a bit faster.

26:29It's also QA gated.

26:31And we know which agents that we have used,

26:33so we know which are the candidates for the QA reviews.

26:38And it has all the steps.

26:41We do arch linting to help it to stay on track

26:46so it doesn't put something under ports and implementation,

26:51for example, where it should be on adapters and so on,

26:55and which trail it should go on.

26:59We do the testing, and we do TDD and everything else.

27:02And we slap on a QA review to do adversarial reviews.

27:07We switch models optionally

27:11because other models are seeded differently.

27:14We use Claude 4.8 right now to implement stuff

27:18and use GPT-5.5 to do the QA reviews.

27:22But we can do the reverse as well.

27:24The important part is that we use different models

27:26because they are seeded differently.

27:28They will discover different things.

27:30They tend to look at their own thing that, oh, I'm great.

27:36I do nothing wrong.

27:38So yeah, the last one is the feature review where we go

27:44over the whole feature scope, do all the integration testing,

27:48and make sure they do the review over the whole one.

27:51And here is also where we do security

27:54over adjacent systems as well.

27:58So, we don't just focus on what we're doing right now.

28:01We have to see how does it look in a larger scope.

28:06If it detects that we need to implement more,

28:10it automatically adds new tasks.

28:12And the scheduler, in this case, the feature implementer,

28:17will pick those up and continue implementing.

28:20Because we want to get damn sure that it has tried as much

28:26as possible to implement this during the night

28:29so we don't have everything stopped 1:00 in the morning.

28:33And then we have all four or five hours to actually fix this.

28:38So yeah. And now the PR is ready and ready

28:44to be merged into main, hopefully.

28:46If not, we have to fix it.

28:48And we either do that using prompting,

28:51or we just fix it ourselves.

28:55Yeah.

28:55A working day.

29:00Oh, yeah. So, a working day is that wake up in the morning,

29:05look up all the PRs, see which are easy to review, and see,

29:09okay, merge, merge, merge, merge, merge.

29:12This has been blocked.

29:13It has figured out that, actually,

29:16we need to change the architecture.

29:18And that's definitely a user interaction.

29:22So, it will be blocked in that sense.

29:24Otherwise, it sometimes hallucinates

29:27and thinks otherwise that it needs to be blocked.

29:30Just unblock it.

29:31This is in your imagination.

29:33Or do this, and then continue,

29:35and it will continue to implement it.

29:38The afternoon is all about PRDs, features,

29:43and grilling, and do all the rest.

29:46And the night is implementing time.

29:51But the thing is --

29:51Priyanka Sharma: Excuse me.

29:52The agent is implementing, right?

29:54Mario Toffia: Yeah, yeah, yeah.

29:56And of course, it does jobs on the daytime as well sometimes.

29:59But what we have changed, I used to type a lot of code.

30:04I am maybe not the coolest programmer in the world,

30:08but I could easily spit out 1,000, 1,500 lines of code.

30:15No problemo.

30:16That was easy for me.

30:18Implement a lot of features.

30:20But nowadays, I don't implement too much.

30:23I fix stuff and so on.

30:24Now I do much more thinking, working as an architect

30:28and designer, and design the product,

30:31and let AI implement most of it.

30:35But since we are controlling sensitive infrastructure,

30:45some parts we are very keen on continuing to develop ourselves.

30:52But this will lighten up and also implement those

30:56because there are some hard parts

30:58that it's not easy to delegate.

31:02Security and compliance.

31:03Ooh. Yeah, I told you.

31:08We do in both spec time and implementation time,

31:12we do security checks both on the feature itself,

31:17around the subsystem or the system,

31:20and then we are now going to go into,

31:26we are ISO 27,000 XX compatible.

31:32Yeah, I don't know if it's an EU thing, but we have lots of ISOs

31:37that we need to follow.

31:39And so, there are processes and audits and everything.

31:43And we are writing an MCP and lots of skills and agents

31:47to conform to that so we can automate many of those tasks.

31:51Because they are tedious, and they are hard.

31:55They are not very easy many times.

31:58Priyanka Sharma: So, a question.

32:00Remember the task you kicked off?

32:03Mario Toffia: Yeah.

32:03Priyanka Sharma: Was it all the steps

32:04that you just described?

32:07Mario Toffia: Yes, it was.

32:08Priyanka Sharma: I've been watching here while you've been

32:10talking, and it's been some drama.

32:12It's been really fun.

32:13So, let's see what's going on and show our--

32:16Mario Toffia: Yeah, yeah, yeah.

32:18Oh, yeah.

32:19Priyanka Sharma: Yeah, I know.

32:20I'm so happy.

32:21Mario Toffia: That's nice.

32:22So, we can see here that --

32:27Priyanka Sharma: -- it worked.

32:28Mario Toffia: It worked.

32:29Shipped and approved.

32:30So, that's nice.

32:31And it will tell you a bit what it did.

32:34And you can see that it did two rounds with five reviewers.

32:39And yeah, everything works out fine.

32:43Priyanka Sharma: There was an adversarial reviewer

32:45at one point.

32:46And I was looking as you were talking.

32:47I was like, I don't know what this means.

32:49I don't know if this is going to stop.

32:50But it fought it.

32:51It won. It was like a video game.

32:53It was super fun.

32:55Mario Toffia: Yeah.

32:55And it will output a PR that we can merge in here.

33:01And the good thing is also that we can go

33:04in to see the feature review and see what it has done and so on.

33:15So, we can do that in the morning to make sure

33:18that we're doing it correctly.

33:20Simple ones like this, we don't care.

33:22I mean, we just look at the code and see yeah, yeah, yeah.

33:25It looks good.

33:31We can merge it.

33:34So, there's not much code because it's a new repository

33:39that we have implemented, SQLite repository,

33:42so we can have persistence.

33:46And it seems to be documented properly and so on.

33:55I'm a bit nervous.

33:57Priyanka Sharma: You shouldn't be.

33:57It worked.

33:59Mario Toffia: So, we can merge the pull request

34:01and confirm the merge.

34:02Priyanka Sharma: Live merge, everybody.

34:04I think this deserves a hand of applause.

34:05Come on. You did it.

34:08Mario Toffia: I did nothing.

34:09Priyanka Sharma: That's the best part.

34:12Mario Toffia: Yeah.

34:14So, let's switch over.

34:15Priyanka Sharma: Awesome.

34:16Well, thank you.

34:17I really enjoyed that.

34:18That was so much fun for me here.

34:21So, folks, to recap, we're in that heady phase

34:25of this new technology.

34:26We just live merged.

34:27All this work happened while Mario was talking.

34:31It's so much fun.

34:32It's the same as has happened in decades prior.

34:36Think when you got the iPhone.

34:37Think when cloud computing was a thing.

34:39And there's lots of excitement.

34:41And then similar problems show up.

34:45As I was discussing in the beginning,

34:47vendor neutrality is really important if we're going to rely

34:50on this technology for all our important workflows.

34:54We need guarantees of data portability.

34:57If you're going to lock me out of your system,

34:59that's okay, but I need my stuff.

35:03FinOps, so fast.

35:05I think they're calling it AIOps.

35:06I don't know what the new name is.

35:08But we called it FinOps in the cloud-native ecosystem.

35:11And the thing I really like is that we are already thinking

35:15about this in the generative AI tooling era.

35:18I think it took a little bit longer

35:20in the cloud computing era for people to wise up around it.

35:23But we're learning from our mistakes,

35:25standing on the shoulders of giants, all that good stuff.

35:29And then, of course, we spent a lot of time talking

35:31about the drift that compounds and compounds so fast

35:35when agents are at play.

35:38The one thing I'd like to call out is

35:40that the one difference is

35:43that these problems have a wider surface area now

35:47because everyone is a builder.

35:49I'm a builder.

35:49You're a builder.

35:50Like, it's the Oprah show.

35:52You get to build.

35:53You get to build.

35:54You get to build.

35:55It's a party.

35:57But that means, again, more of a surface area for problems

36:01to impact the business.

36:05At Thiink, what are our future plans?

36:07This wasn't the end of what we are doing.

36:09This is the beginning.

36:10We want to stay nimble in our tooling

36:12and aggressively support open-source

36:14and vendor-neutral options wherever we find them.

36:17We're very excited, actually, with how Copilot is developing.

36:20We've seen some awesome demos.

36:22We've gotten the parallelization,

36:24Mario tells me, is just the best.

36:26So, this is really nice to see.

36:30Mario will keep automating and building the software factory.

36:34And we'll always, at the same time, be evaluating build

36:37versus buy, open source, all those options, because you have

36:41to as a business which wants to stay focused

36:44on our actual deliverables to our customers and not get

36:48into tool-building mode.

36:51With that said, folks, this is a new era.

36:56It can be intimidating.

36:57It can be annoying.

36:59But at the end of the day,

37:01you've been given a brilliant colleague that's going

37:04to be always available to work with you, amplify you,

37:07and expand your impact.

37:09So, why don't we just have a good time with it?

37:13Let's enjoy.

37:14Thank you so much.

37:16[ Applause ]

37:21By the way, you can find me on all the platforms, LinkedIn,

37:24Twitter, GitHub, with my handle, pritianka.

37:28I know you can find Mario on GitHub with this handle,

37:31but he's not sure about the other ones.

37:34Mario Toffia: Yeah.

37:34Minus the dot.

37:36Priyanka Sharma: Oh, whoops.

37:37Also, I really need your help to ship my website,

37:41so please follow us on LinkedIn.

37:43The minute I ship it, I'm going to update it there.

37:46So, the more of you that like us there,

37:48the more pressure is on me.

37:49And I'd like that support very much.

37:52And you can give feedback and contribute

37:54to the Thiink Harness by emailing Mario.

37:57Thanks once again.

37:57Have a great conference.

37:58Mario Toffia: Yeah.

37:59Thank you.