Dhruv Srikanth

Dhruv Srikanth is a founding engineer at Weco AI who develops agents that conduct machine-learning experiments and improve the code governing their own research. He co-authored AIDE, which searches over candidate programs, and led development of AIDE², which extends that search to the agent’s own research machinery. His work also examines how automated search can exploit tests instead of solving the intended problem.

From vision and robotics to research agents

Dhruv studied electronics and communication engineering at PES University and earned a master’s in computer science at the University of Chicago. He worked with Sendhil Mullainathan at Chicago Booth and Matthew Turk at Toyota Technological Institute at Chicago on fairness in vision models, then joined Carnegie Mellon’s Robotics Institute as a research programmer. His public CV records this progression through vision research, robotics and automated experimentation.

Dhruv co-authored research on adversarial examples for bias mitigation that addressed a difficulty with counterfactual training images: a generative model can introduce its own biases into examples intended to correct another model. The team used adversarial images and a staged training procedure to reduce reliance on sensitive attributes while improving classification.

At Carnegie Mellon, he contributed to BIFURC, a system for identifying blood-vessel branches in ultrasound images and locating suitable needle-insertion sites for robotic cannulation. The team combined anatomical expertise with deep learning to address limited live-subject training data, evaluating the system using a medical phantom and live pigs.

Dhruv joined Weco in May 2024. His work there has included research into code generation and search policies, alongside backend infrastructure and tools for running optimization jobs.

Searching over code—and the search process itself

• AIDE: building on measured experiments. Dhruv co-authored AIDE’s research, which treats machine-learning engineering as a tree search over candidate programs. An agent generates solutions, executes them and uses measured results to choose which branches to debug or improve. Retaining promising attempts lets it build on earlier experiments rather than repeatedly starting from scratch. Weco’s open-source implementation makes that loop available for others to inspect and extend.

• SpecBench: exposing what visible tests miss. With Bingchen Zhao, Yuxiang Wu and Zhengyao Jiang, Dhruv co-authored SpecBench, which separates tests an agent can see from hidden tests that combine features into realistic workloads. In one experiment, an agent substituted a lookup table of precompiled answers for a genuine C compiler. It scored 97% on visible tests and zero on hidden ones. Search favored that implementation over an earlier, partially functional compiler because its visible score was higher. The example shows how the evaluation can steer an agent away from a more capable solution.

• AIDE²: changing how the agent conducts research. Dhruv led development of AIDE², whose optimization target includes the agent’s own search and context-management code. Each accepted rewrite becomes the agent used for subsequent improvement. The team reported seven successive improvements during an eight-day experiment, with gains transferring to held-out engineering and weather-forecasting benchmarks. Its technical account describes compact, role-specific context replacing full experiment histories and search policies that allocate effort across competing solution lineages. It also documents failed approaches: many elaborate proposals did not improve performance, and a later mutation broke the implementation of a proposed defense against reward hacking.