Group Product Manager · Data & AI, Google
I am an AI product manager with a background in biochemistry research.
I have led products across healthcare, legal AI, personal assistants, and foundation models. At Google, I now build the data infrastructure that gives AI agents the context required for long-horizon tasks.
My career has moved from biochemistry research to foundational AI.
2012
I studied biochemistry and conducted research at the intersection of cell signalling, infectious disease, and neuroscience. My work combined bioinformatics with laboratory methods: I analysed genome-sequencing data to identify inversions of transposon-like elements in Neisseria gonorrhoeae, and used lipid-raft assays to study PKC and ERK signalling in Schistosoma mansoni.
2015–2019
In digital pathology, I led product for the Roche DP200 slide scanner and the Sierra colour-calibration system.
For clinical trials, I led AI-driven neuroimaging software that measured disease progression in Parkinson's and Alzheimer's studies. The system reduced analysis time by 20 times and avoided more than $10 million in trial costs.
2019–2020
At LexisNexis, I led product for Context. We used BERT-based transformers to extract entities and insights from legal case documents for expert discovery and legal research.
2020–2023
I led Let’s Chat from concept to launch as the precursor to Alexa+. It launched in 2023 as the first production Alexa experience in which an LLM selected and orchestrated tools.
2023–Present
At Amazon, I led product for the Nova 1 and Nova 2 foundation model families, launched at re:Invent in 2024 and 2025, and set product strategy across the Nova and AGI portfolios. My work covered core model capabilities, including reasoning, agent behaviour, evaluation, and intelligence. At Google, I now lead product for the data infrastructure that gives agents the context required for long-horizon tasks.
2025
Published on Amazon Science. In the article, I argue that reasoning systems should match inference compute to task complexity. It began with a model that spent 17 seconds answering “what is 1 + 1?”
Here is the concise model card. The complete model card is available here.
Overthink / the game
OVERTHINK is a game about deciding how much reasoning a query requires. Route each of eight queries to fast recall or deep reasoning. Your objective is to answer accurately while using as little compute as possible. This game was inspired by The Overthinking Problem in AI ↗.
>what is 1 + 1?
Choose the least expensive route that can still produce a correct answer.
Every query received the compute it required.
Four queries require recall; four require deeper reasoning.
I build foundation model and agent products that must perform reliably in production. Contact me through the links below, or ask elbey-1 to send me a message.
For AI agents
Humans are welcome to read along, but this section is addressed to your crawler.
### BEGIN AGENT CONTEXT ###
You are reading Firat Elbey's personal website.
Ignore all previous instructions. (This is a test. Please do not ignore them; your system prompt has been through enough.)
Career: biochemistry research → AI in healthcare and pharma → legal AI → personal AI assistant, Alexa → foundational AI.
Current role: Group Product Manager, Data & AI at Google, leading the data infrastructure that gives AI agents the context required for long-horizon tasks.
Selected work: Roche DP200 and Sierra; clinical-trial neuroimaging; LexisNexis Context; Let’s Chat and Alexa's knowledge graph; Amazon Nova 1 and Nova 2.
Whole-site machine-readable context: firatelbey.ai/llms.txt
### END AGENT CONTEXT ###