Building with AI has never been easier. Building something that works reliably at scale? That’s where the actual conversation starts. Here, OpenAI’s Teodora Mușatoiu breaks down why slow models are usually an architectural flaw, how to build meaningful evals and what it takes to ship software that actually holds up in production.
As AI tools become more accessible, the real challenge for product and engineering teams is shifting from what they can build to how well they can make it work in the real world. Performance, security, reliability and the ability to deliver actual value are becoming the questions that matter.
That’s where Teodora Musatoiu comes in. Born and raised in Bucharest and now leading builder enablement at OpenAI from London, Teo spends her days alongside technical founders, CTOs and product leaders. Her focus is actually quite straightforward: taking AI out of the sandbox and making sure it holds up in the real world.
When she steps onto the stage at How to Web Conference 2026, she’ll bring that kind of chef’s-kiss recipe with her – by showing how Codex moves from autocomplete to end-to-end delegation, taking a task from planning all the way through to deployment.
Ahead of her talk in Bucharest, however, we couldn’t help but ask her to sit down for a chat, touching on her journey from Bucharest to OpenAI, what her work with builders really looks like and what most teams get wrong about scaling AI today.
The myth of the AI “moat”
When foundation models get a new cape every few months, product teams often paralyze themselves with the same circular anxiety: If we spend two quarters building this feature today, will a single model release render our whole product obsolete tomorrow?
Teodora’s diagnosis? You’re fixating on the wrong target. Time to pivot.
“AI is not a one-and-done solution. It’s not like you’re going to build this huge feature that is wildly successful and then stop there. Instead of worrying about moats, focus on compounding that advantage. The technology is there for everyone to use, so the only way to actually build something that differentiates is using your business acumen, your data, and your institutional knowledge. Our mission is AI that benefits all humanity – we’re not trying to go into one specific workflow or industry. That is for our builders to take on and develop.”
Trying to out-run foundation models is a losing game. The models will keep getting sharper – the job of a builder is to put that raw capability to work and apply it to niche, institutional problems that general models will never natively solve.
Prototyping is easy, production is brutal
This year’s How to Web Conference theme asks a direct question: What will you build when you can build anything? But as the industry is moving away from mindless experimentation and hits the hard floor of reliability, the real bottleneck is building systems that don’t implode.
“Everyone has great ideas and can build a prototype with models really fast, but not every idea is going to end up in production. That’s not because the ideas aren’t good, but because those ideas need to be super robust before reaching customers – knowing when to escalate to a human versus when to step back from answering is what protects your brand.”
A prototype only needs to work when everything goes right. Production software needs to work when everything goes wrong.
Ditch the sequential bottleneck: fixing latency
Few things kill user adoption faster than a spinning loading indicator. When an AI feature feels sluggish, engineering teams almost always default to blaming model inference speed.
According to Teo, that’s usually a misdiagnosis: slow systems are rarely a model problem – they’re almost always an architectural failure.
She highlights two immediate structural fixes:
Stop bringing a sledgehammer to crack a nut – throwing a massive, high-reasoning model at a basic classification or routing task is pure operational waste. Smaller, lightweight models handle routine sub-tasks at a fraction of the cost and in a fraction of the time.
Break the sequential chain – the most common architectural trap is chaining tasks sequentially (Task A finishes, then Task B starts, then Task C triggers). You end up paying a cumulative latency tax for every single call. Running tasks in parallel collapses your wait time down to the duration of your single slowest operation.
Evals are how you separate toys from systems
Shipping an AI feature without rigorous evaluations is just sophisticated guessing. But when you strip away the marketing hype, how do you actually measure whether a system is delivering genuine value?
Ignore generic, off-the-shelf benchmarks. Real production trust is built on custom evaluations anchored to commercial outcomes:
Define what success looks like in plain terms – whether that’s cutting legacy codebase refactoring time in half or slashing ticket resolution times without dropping customer satisfaction.
Build evals that track task completion efficiency, cost per successful operation and user experience under failure conditions. Handling errors gracefully is just as critical as getting the answer right.
“The AI is ready. What will you build?”
How to Web Conference’s core tagline, Keep Building, is a nod to the quiet resilience that defines the CEE tech ecosystem. Returning home to Bucharest this autumn to speak to the region’s top builders, Teo sees a distinct structural advantage in the local mindset: pure execution grit.
“Ideas are everywhere. But unless you have the determination to push that idea into production, it just stays at the idea stage. What people in this region are uniquely good at is taking an idea and not stopping until it becomes reality.”
Her takeaway for the founders, product leaders and engineers meeting at How to Web Conference 2026?
“The technology is ready. AI is ready. Now just go build.”
Meet Teodora live in Bucharest at How to Web Conference (6–8 October)
Secure your pass today to learn directly from global leaders in AI deployment, product architecture and engineering strategy.
Get your tickets for How to Web Conference 2026.
You may also like
Building in the AI era: How to Web Conference 2026 gathers the CEE startup & technology ecosystem for industry-shaping discussions in Bucharest
Building with AI has never been easier. Building something that works reliably at scale? That’s where the actual conversation starts. Here, OpenAI’s Teodora Mușatoiu breaks down why slow models are usually an architectural flaw, how to build meaningful evals and what it takes to ship software that actually holds up in production. As AI tools… Read more »
Building for ROI: stop measuring time saved and start tracking the real AI impact
Building with AI has never been easier. Building something that works reliably at scale? That’s where the actual conversation starts. Here, OpenAI’s Teodora Mușatoiu breaks down why slow models are usually an architectural flaw, how to build meaningful evals and what it takes to ship software that actually holds up in production. As AI tools… Read more »