Pratyusha Duvvuri

Pratyusha Duvvuri is an engineer and founder who helped build Galileo’s Agent Graph. In 2025, while leading work on agent observability and distributed tracing at Galileo, Pratyusha saw customers reconstructing an agent’s actions by hand from logs and traces. She and a collaborator prototyped a graph of those interactions and released an early version in a week. In one support-ticket example, the agent picked the right tools for the job. The graph showed why it still failed: it tried to escalate the ticket before it had created it. Pratyusha could follow that sequence back to the code that ordered the calls.

Pratyusha had run a startup as a student at Carnegie Mellon. When investors encouraged her to leave school, she chose to stay for the degree and the college experience. She also doubted she wanted to spend five years on that particular idea. At Carnegie Mellon, she built a Game Boy emulator with Tess Chan and Adolfo Victoria. She designed the controller driver and worked on communication between the processor and the programmable hardware that ran the games. Before joining Galileo, she worked as a quantitative engineer at Two Sigma, where she helped launch investment portfolios.

Who pays for an AI workload?

After Cisco acquired Galileo, Pratyusha stayed on as technical lead for tokenomics. Her team spoke with roughly 100 companies about how they tracked AI spending. Even calculating the bill required more than counting tokens. Customers negotiated different rates, and providers sometimes revised charges after reporting them. Assigning the bill raised a separate question. An engineering manager wanted to see which people on a team were using a tool, while a financial analyst needed to charge the right cost center. That cost center could span several teams. Pratyusha’s team mapped the same usage to both reporting structures and took the product from its first customer conversations to general availability in ten weeks.

Pratyusha weighs model cost against what a better result is worth for the particular job. On an internal task, her team kept the cheaper model after a more expensive one raised the evaluation score from 96% to 99%. She would pay for that gain on a task where customers would benefit from the improvement. For back-office work, she proposes comparing people and agents by the cost of a correctly completed task. Both should do the same work to the same quality standard, and the cost on each side should include retries, corrections and human supervision.

Returning to founding

In October 2026, after a year and a half at Galileo and five months working with Splunk and Cisco colleagues after the acquisition, Pratyusha announced that she had left to start another company and was looking for co-founders. She had promised her teammates she would stay through the release.

One question she investigated was how a distributor decides whether to ship another order to a retailer who still owes money. Across 50 conversations about supplier credit, focused on the United Arab Emirates, she found that the same overdue invoice could call for different responses. A missing delivery receipt might hold up payment even when a retailer had the cash. A retailer waiting for its own customer to pay posed a different risk. Pratyusha saw an opportunity to help credit teams decide whether to release the next order. Who would pay for that help, what data they could share and whether a product would improve the decision remained open questions.