Devin Gupta

Devin Gupta is a member of the technical staff at Parallel Web Systems who leads work on the Parallel Decisions API. The API turns questions about the live web into structured answers with citations. A team evaluating a vendor, for example, can ask whether it supports single sign-on alongside other questions about the same company. The questions can share the retrieved evidence, while each answer goes into a separate field that software can use and a person can check.

From investment research to web agents

Devin studied physics at Stanford. In 2023, he built a workflow at Duco Experts to screen policy headlines and test prompts against human labels. At Atomic Vaults that fall, he analyzed exchange data around stock splits to look for retail demand for fractional options. During a summer in long/short equity research at Balyasny Asset Management in 2024, he investigated semiconductor supply chains through expert interviews and financial modeling and pitched short positions. He also brought outside datasets into the firm’s research platform and used its embedding models to map related material.

At Distyl AI in summer 2025, Devin co-developed a system that used web search to generate strategic proposals for large companies. New findings fed into another round of reasoning and search. He separately led work on hierarchical planning for a browser agent, breaking larger tasks into smaller steps and testing the approach on the shopping administration portion of WebArena. At Parallel, he adapts open models for research agents and works on how the models and the software that runs their tools and search loops are designed together. He has also helped integrate data from partners into Parallel’s research systems, giving agents access to evidence beyond the public web.

The cost of finding evidence

Devin and Leela Karumbunathan co-authored the introduction to PWS-Decisions, a benchmark that tests whether a system can find a specific fact on the web and use it to choose an answer. Parallel released 100 questions through its Hugging Face dataset, which requires sign-in, and kept another 200 private. Their cost argument starts with the evidence: a classifier can be cheap to run once it has the relevant facts, but a separate lookup for every answer can erase much of that advantage. Devin’s AI Engineer Code 2026 session abstract puts the difference in scale at 104 tokens for the question versus 4,777 for the evidence on PWS-Decisions. Reusing evidence across questions is one way the Decisions API addresses that cost.

For longer assignments, the choice of query can matter as much as whether evidence is reused. In Parallel’s test of a labor market research assignment, Devin pointed to how the agent gathered statistics across several states in fewer steps by issuing more focused queries. In that case, the savings came from covering more ground per search.