Machine learning intern · Costa Mesa, CA

Production demand forecasting for McDonald's, Taco Bell, Jack in the Box and Chipotle — thousands of locations, daily horizon. I own the model lifecycle end to end: 30+ engineered features drawn from weather, zip-code demographics, holidays, school calendars, national events and macroeconomic indicators; cross-validated tuning; back-testing across more than two years of history.

Forecast error came down from 12% to 5%. Downstream that is tighter labor scheduling and lower food cost, which is the entire reason the model exists.

I've run this alongside a full Berkeley course load in two consecutive summers — CS 61C in 2025, CS 184 and CS 161 in 2026.

Residual Robotics · 5-person team · Berkeley

Robot policies don't generalize the way language models do, and a large part of the reason is data: the manipulation corpora that exist are small and narrow relative to what the problem needs. We're collecting mass UMI data from skilled laborers in India to widen that, and selling it to robotics foundation model labs. The hardware is a handheld gripper you can carry into any environment and record real manipulation with, no robot required.

I design the gripper in CAD — full assembly in Onshape, jaw linkage, camera mount, IMU housing, fiducial plate. I re-specced it off its $298 GoPro toward a webcam, microcontroller and IMU, then did the design-for-manufacturing work to get all 16 printed parts off the bed with zero slicer supports.

CS 184 final project · 2-person team with Ninad Atale · in progress

Commercial FEM solvers make you re-solve from scratch for every new load case. For linear-elastic materials that is wasteful, because any load combination is a superposition of load cases you have already solved. We are building a solver that generates its own training set by sweeping load cases automatically, training a graph neural network on the mesh, and exposing the result as a viewer you can push on and watch deform in real time.

Three pieces. A mesher: Assimp to load CAD and COLLADA, a manifold validity check, TetGen to discretize into tetrahedra. An evaluator that applies fixed supports and point loads and solves the elastostatic PDE for displacement and stress. Then a GNN over the mesh graph, message passing on node and edge features, targeting a 10-minute training budget on meshes under 10,000 elements.

Scoped to linear elasticity and direct forces; no torques, no anisotropy. Accuracy is measured against our own solver's solution, not against a commercial one, because matching topology across representations is its own problem. Proposal submitted 27 Jul 2026. Nothing here is a result yet.