So hi, everyone. I'm Rashi Agrawal, and I lead engineering teams building powerful AI products and solutions at Good Leap.Good Leap is a technology company delivering best in class financing and software products for sustainable solutions. I'm super excited to be here at the Women in Tech Global Conference. Today, I wanna take you behind the scenes of what it really takes to build and scale AI in a company. Not just the hype, but the hard work, the strategy, the team structure, and, yes, the bumps along the road. Let's start with what most people see when they look at AI today. From the outside, AI can look like magic. You type a prompt. It responds with something clever. Executives love the demos. The UI looks sleek. The wow factor is real. But what most people don't see is what it actually takes to make that work at scale.

It's a full stack engineering beast. Data pipelines, evaluations, security, latency constraints. That simple chatbot that you see might be hiding a team of 15 engineers and a platform stack seven layers deep. So today, I'll walk you through what's beneath that surface. If you are trying to do AI at scale, this talk is for you. We have already entered the era where real transformation is happening at the top. Your CEO, your board, they are already thinking about how AI can be a strategic lever for business transformation. It's no longer a matter of if, but how fast and how effectively. But here's the catch. The thing to get right isn't the model. It's the mindset.

And that mindset shift means we start not with the tech tools, but with the business problems. Ask, what are your biggest challenges? And not, can we use chat GPT for this? That reframing changes everything. One of the most valuable steps we took early on was an AI fitness check. Do we have the right data, the infrastructure, the talent, the risk appetite. That honest self assessment brings clarity, and it helps to set realistic expectations of what's possible. But readiness alone isn't enough. Let me give you an example of what happens when you skip a critical foundational step. In the early days of our AI journey, we built an AI powered conversation review tool, and it addressed a very real business problem, helping compliance teams flag risky conversations and relieve the burden of manual review. The prototype worked. Technically, it was strong. The data was reliable. The infra was solid, and the use case was clear. But we skipped a important layer, executive buy in.

There was no strategic sponsorship, no clear push to move it beyond the pilot phase. And without that top down momentum, it quietly faded out. No traction, no adoption. The takeaway. Even when you've identified the right business problems and done your AI homework, skipping any layer of the foundation, especially leadership alignment, means AI doesn't scale. Prototypes are great for exploration, but we are way beyond the experimentation phase now. Real transformation takes commitment, and it starts at the top. Next, let's talk about what powers AI at scale and what's inside the engine room. The part, teams. We learned early on that traditional hiring will slow us down. The market is competitive. Time lines are tight. And let's be real. AI powered interview cheating is a thing now. That's like a whole new can of worms. Resumes look great. Interviews go smoothly. But you're left asking, did we just hire a genius or someone who let ChatGPT ace the interview for them?

So instead of relying solely on traditional hiring, we augmented our team with a mix of strong engineers and trusted contractors. That blended model has helped us move rapidly from prototype to production without getting bottlenecked by talent gaps. But the magic isn't just in hiring speed. It's in cross functional collaboration. We brought together PMs, machine learning and back end engineers, and UX designers, people who really understand human AI interaction and can turn smart models into intuitive experiences. Equally crucial is their awareness that these systems are nondeterministic, meaning they won't behave the same way every time. So we design for ambiguity, guide expectations, and build trust. And, of course, we constantly think about build versus buy.

We continuously ask ourselves, where should we go deep, and where can we just plug in something solid that already exists? Because let's be honest. You don't want a team of engineers spending months building something something that you could have just subscribed to for $200 a month. Our goal is simple. Maximize ROI, but not over engineer. Now let's switch gears to the tech side of the engine room. At Goodleap, we treat AI not as a model you plug in, but as a system you architect. Success depends on how well your entire pipeline works and not just on how well engineered your prompts are. Our architecture reflects that. It spans multiple layers from raw compute, data pipelines, to model logic, knowledge retrieval, and user facing interfaces.

Each layer from infrastructure to experience has to work in harmony for AI to be scalable, safe, and effective. And, yes, each layer serves a critical function. Skip one, and you don't just get a weaker product. You get instability at scale. We learned our lessons the hard way that if you don't test for latency or build in evaluations early, it will come back to bite you in production. Finally, we stay model agnostic by design. Today's best model might be tomorrow's best. Flexibility gives us leverage, especially as pricing, latency, and licensing continue to shift fast. Sometimes, the most strategic thing that you can do is not commit too early. Now I wish I could tell you it all went smoothly, but, of course, it didn't. Here are the four biggest bumps that we hit. Privacy. If you don't plan for it upfront, your launch will stall. You need privacy preserving design from day one.

Latency and cost, these are real enterprise killers. A thirty second response time or a surprise 80,000 token $80,000 token bill? Not fun. Team alignment. Most pilots don't fail because of tech. They fail because of misalignment between PMs, engineering teams, business teams not working as one. Enterprise and risk. AI isn't just code. It's bias, ethics, compliance. You need people who truly get that full picture. Now that we build something, the real question is, how do we get people to use it? Because success isn't about shipping AI. It's actually about adoption at scale, and that takes a shift in how people work, trust, and engage with AI. Let's break it down into three pieces, enablement, governance, and integration. What good is a powerful AI tool if no one knows how to use it or worse, is too intimidated to even try?

That's why our mantra has been make AI accessible. We focused on empowering teams across the company and not just the engineering org. We created adoption channels through AI gills, training sessions, and lightweight documentation. And to keep things fun and experimental, we hosted hackathons, a great way to surface use cases, foster creativity, and build grassroots excitement around AI. Now imagine, what's more terrifying than an AI tool gone rogue? One that fails an audit? As much as we wanted fast adoption, we knew we needed guardrails. So we built a governance framework designed to scale with the org without slowing teams down. That meant weaving in legal and security reviews, compliance by design, prompt versioning, and audit readiness from the start.

Think of it as AI DevOps meets policy, automated, trackable, and sane. Of course, even the safest system won't help if no one actually uses it. Do people really want another tool? No. They just want their current tools to be smarter. So instead of launching new portals or flashy dashboards, we focused on embedding AI into existing workflows. We asked, where are people already working, and how can we make those tools more intelligent? The best feedback we got was, I didn't even realize it was AI. That's the goal. When adoption feels natural, seamless, and useful, people stop resisting it and start requesting more of it. So all the strategies, team structures, and culture controls are great. But at the end of the day, how are we truly transforming the business?

Because if AI isn't delivering measurable value, it's just another shiny new toy. Here's where the rubber meets the road. Did our AI products actually work? We didn't settle for vague goals like get better answers or improve user experience. We aimed for outcomes like reduce support tickets by 30%, cut processing time by 40%. And to track real impact, we measure across three levels. Business, what has changed for the user or the bottom line? Model, is it accurate, safe, and hallucination free? System, what's the latency, cost, and uptime? And here's the part I can't stress enough. Evals, evals, evals. Lack of evaluation is a direct path to hallucinations. If you're not actively measuring what your model is saying and how it's behaving in real use cases, you're just hoping it gets it right. Hope is not a strategy. Evals help you build trust, spot issues early, and guide continuous improvement. We started small, manual spot checks, basic scoring, and built from there.

Each iteration made the system smarter and safer. It's never perfect, but it's always getting better. And that's what progress at scale looks like. If you take away just one thing today, let it be this. AI is a muscle. You don't get six pack abs from just one sit up. Same with AI. You've got to build it. If this were a cheat sheet, I'd say start small, but tie it to real business problems. Build a flexible, modular foundation. Don't marry one model or vendor. Empower your teams. AI should amplify humans, not replace them. Let's move beyond the hype and start solving the right problems at scale. Thank you all for joining me. Let's connect. I'd love to hear what you're building, what challenges you're facing, or what resonated today. You can find me on LinkedIn. And, yes, Goodleap is hiring. At Goodleap, we are not just building tech for the sake of it. We are using it to power meaningful change.

Through our nonprofit partner, GivePower, we've improved the lives of over 2,000,000 people across 28 countries, bringing clean water, solar power, and real infrastructure to communities that need it most. Whether it's building smarter AI or helping bring electricity to a school across the globe, our work has purpose, and we're just getting started. Let's build the future together. Thank you.