Using AI Tools as a Scientist in Tech

While ChatGPT has been around since 2022, it’s only in the last year and a half that I’ve seen an acceleration in the adoption of AI tools at my workplace. In this post, I wanted to document the evolution of what tools I’ve encountered and used in my role as an applied scientist at Amazon. It’s by no means a recommendation for the best way to adopt these tools or what the best tools are, only a reflection of my personal experience.

The Olden Times Pre-2024, I wrote almost every line of code manually. Even before ChatGPT was released in 2022, however, there were some increasingly advanced autocomplete coding tools like GitHub Copilot. Later, Amazon came out with a similar product called CodeWhisperer. While I experimented with these autocomplete tools, I personally never really got into using them because I didn’t feel like they significantly improved my productivity and were sometimes more of a nuisance than anything.

Cedric The end of 2024 was when the way I started coding started changing significantly. Around this time, we got an internal ChatGPT-like tool called Cedric. With Cedric, I soon learned that AI can write code snippets much faster than I can, with well-formatted and almost invariably superior syntax. The more I used it, the more I trusted the code it wrote. This started a snowball effect where I wrote less and less code manually as the year progressed.

Cline Around June 2025, one of my teammates told us about a Visual Studio Code extension called Cline. We also learned about MCPs (Model Context Protocols), which allowed various other services to connect to AI. This was a major breakthrough because up until then, AI was just a chatbot siloed from our internal systems, limiting its usefulness. With MCPs, the AI could now do things like run SQL queries on Redshift clusters or read our internal documentation and code packages. However, I never really adopted Cline as part of my workflow, mainly because it was pretty unreliable – it would, for example, frequently reach token limits and just stall out. But the potential it represented was different from anything I’d seen before. One of the things I tried to get it to do was to “look” at an internal metrics dashboard and check for any anomalies. I remember being really impressed watching it control my browser, move the cursor, and click through various dropdown menus without any additional guidance.

Kiro Later in the summer of 2025, Amazon released an AI-centric IDE called Kiro. Kiro could basically do everything that Cline did in a more stable way. However, I didn’t adopt Kiro as part of my daily workflow until Kiro became more integrated with SageMaker in 2026. By 2026, nearly all of the code I wrote was generated first with AI. With Kiro, I also switched from writing code primarily in Jupyter notebooks to Python scripts with cell markers. This was because the AI had a much easier time editing .py scripts, which only contained code, instead of .ipynb files, which were a mix of both code and output.

I also started using Kiro as a personal assistant of sorts by having it synthesize my notes, Slack messages, e-mails, and calendar. The use of AI this way sometimes led to hilarious situations where I realize I’m using AI to summarize a newsletter that itself was created by someone else using AI on top of notes that were written by AI… it’s AI all the way down.

Claude Code and more The AI landscape moves fast – half a year after Kiro was released, there were already growing complaints at the company that we didn’t have access to Claude Code, which was quickly becoming the industry standard. Recently, Amazon rolled out Claude Code and Codex to its employees and I expect this will again bring new changes to our workflow.