Sravan
Sr. Staff Agentic AI Software Engineer at Google
Optimizing Agentic AI workflows and hardware-software co-design for TPUs at Google Cloud.
Intro
I'm Sravan, a Sr. Staff Agentic AI Software Engineer at Google. My career has been defined by building and scaling large-scale AI systems, from leading the science teams behind Amazon Rufus and Nova Foundation Models to my current focus on Agentic AI and TPU optimization at Google Cloud. I hold a Dual Degree from IIT Madras and have spent over a decade pushing the boundaries of NLP, ASR, and Generative AI. I am deeply passionate about the economics of inference and the necessity of hardware-software co-design to make AI sustainable and accessible. Beyond my core engineering work, I enjoy mentoring startups through the Google for Startups Accelerator and helping founders find top-tier AI talent. I'm always looking to connect with people who are moving past the hype to solve real-world problems in token economics, sovereign AI, and autonomous agent workflows.
Networking
What I can offer
- ›Startup mentorship via Google for Startups Accelerator
- ›Deep expertise in LLM post-training and inference optimization
- ›Guidance on scaling global science and engineering teams
- ›Technical strategy for conversational AI and foundation models
What I'm looking for
- ›expanding my professional network
- ›exploring mutual opportunities in Agentic AI and LLM infrastructure
Best fit for
Focus
Current interests
Core competencies
Background
Career
Progressed from software internships to Applied Scientist roles at Amazon, eventually leading science teams for Rufus and Nova Foundation Models before joining Google Cloud/BigQuery.
Education
Dual Degree (B.Tech + M.Tech) in Computer Science from the Indian Institute of Technology (IIT), Madras.
Achievements
- ›Led development and launch of Amazon Rufus conversational shopping assistant
- ›Delivered 1M context support for Amazon Nova LLMs
- ›Authored 50+ patents (pending USPTO approval)
- ›Scaled AWS LLM/ASR science team from 2 to 30+ members
- ›Achieved 3.7x improvement in Compute Carbon Intensity for Google Ironwood TPU
Opinions
- LLMs currently memorize and retrieve rather than understanding reality or physical mental models.
- Auto-generated context files like AGENTS.md often hurt performance due to bloat and unnecessary reasoning token consumption.
- Lowering token unit economics is the most effective strategy for AI adoption in price-sensitive markets.
Personality
Communication style
Professional, analytical, and highly technical, often using bullet points and bold text for emphasis.
Formality — 8/10
Vocabulary