Skip to main content

Curiosity, opportunity, and following reinforcement learning wherever it leads

Kiro Principal Applied Scientist Sahika G. on following her passions and never running out of frontier at Amazon.

Applied scientists at Amazon can follow many paths

Sahika G. was not trained to build things. She was trained to prove them. As a theoretician at GE Global Research, she worked on edge devices, healthcare, aviation, and power grids. Precise environments where the machines were built by people who built jet engines.

Then machine learning started moving fast, and former colleagues at Amazon kept calling.

“At GE, I was improving the performance of existing machines,” says Sahika. “At Amazon, the scale was entirely different. It moved from high performance to high volume.”

She joined about eight and a half years ago. Her first project? Teaching toy cars to drive themselves using reinforcement learning on a platform called AWS DeepRacer.

“We learned that toy cars are not built by people who are building jet engines,” she notes. “Things can go wrong all the time. And those were the things that I learned—and how to deal with it.”

Where curiosity took her

From AWS DeepRacer, she moved into a whole-page optimization for Amazon search. From search, she moved to Rufus, Amazon’s conversational shopping layer (now Alexa for Shopping), where the same techniques she’d used to steer robots could steer a large language model away from hallucinations. And from Rufus, she moved to Kiro, the agentic AI platform for software development.

Each time she had the opportunity to explore something new, the domain changed completely. The discipline didn’t.

“I like it because you’re learning something new every time,” she says. “If you’re with a company for a long time, you need that excitement to keep going. Amazon is a great place to provide that opportunity: You can move across different teams and it’s relatively flexible.”

Each move was catalyzed by a conversation with a senior principal scientist who pointed her toward the next problem worth solving.

How ideas take shape

Today on the Kiro Science team, she’s applying reinforcement learning to push the boundaries of agentic AI for coding. The work happens the way most new ideas happen at Amazon: small groups, fast iteration, and honest feedback.

“A new idea at Amazon starts with a small team building probably a pet project that they are really proud of,” she says, “and then it just takes on this life from there.”

She encourages her team to keep those pet projects alive and keep the spark lit. “If you don’t have that spark in your fire pit, you won’t be able to heat your home.”

The unexpected advantage

The circuitous route Sahika took to Amazon—theoretician to robotics scientist to search optimizer to conversational AI builder to agentic coder—turns out to be the point.

“I’m not a builder from my training,” she says. “But I really enjoyed that you write your theorems, your proofs, and then you take it to the next level. What can I do with this, and how can I help the community?”

At Amazon, she found out. And keeps finding out.

“There’s always an opportunity to work on the next frontier. I’m pretty sure that Amazon will be at the forefront of what’s coming next.”

"There's always an opportunity to work on the next frontier. I'm pretty sure that Amazon will be at the forefront of what's coming next.”