Rohan Pandey

Rohan Pandey is a software engineer on DigitalOcean’s dedicated inference team, where he measures how much traffic a language model can serve on a given GPU while meeting response-time targets. His benchmarks help the team decide how much capacity to provision and where to route requests. He is also co-founder and CEO of Qorbit Systems, which builds recovery hardware and flight-recording software for model rockets.

From student rocketry to Qorbit

At the University of Washington, Rohan led engineering on ARES, a student project that combined GPS-guided rocket recovery with software for recording and transmitting flight data. In December 2025, he co-founded Qorbit with Rohit Pandey, who leads hardware development. Qorbit helps competition teams check their recovery systems before launch, is flight-testing a backup for cases when the primary system fails to deploy, and gives teams a flight record they can share and inspect after landing.

Alongside his studies in applied and computational mathematical sciences, Rohan modeled interactions between CAR T cells and tumor cells at Fred Hutch using data from patients with leukemia. At PwC he built Python data pipelines, and at Mercor he trained and evaluated sentiment-analysis models on domain-specific text. In March 2026, he completed his bachelor’s degree, joined DigitalOcean and founded Lumina Research, whose Research Studio helps researchers prepare papers for submission. He continued at Washington as a master’s student in applied mathematics.

Measuring language-model performance

At DigitalOcean, Rohan built an agent that tunes GPU kernels. With Balaji Varadarajan, he also presented the team’s vLLM serving tuner. Rohan demonstrated how it uses Optuna to choose which settings to try next based on throughput measured under latency targets. A short chat and a batch of classification requests place different demands on the same server. In one nine-trial run on a distilled DeepSeek-R1 model, the best configuration doubled token throughput and nearly halved the wait between successive tokens in a streaming response. In another study the presenters described, an apparent 4% gain fell within ordinary variation between runs.

Searching for smaller circuits and training for shorter answers

At Washington’s Math AI Lab, Rohan led reinforcement-learning experiments on finding the fewest additions and multiplications needed to compute a polynomial. He and his colleagues had studied how to build a circuit one operation at a time in CircuitBuilder. In FactorLibrary, Rohan, Michael Ruofan Zeng and their co-authors also worked backward from the target polynomial. The agent chooses how to split an expression into a sum. A factorization algorithm breaks the pieces down further, and the library saves useful expressions as intermediate goals for later attempts. This reduces the number of decisions the agent must learn. On their small two-variable problems, a top-down agent using tree search gave the most stable results. A version trained without tree search reached a slightly higher peak in about one-fifth of the training time, but its performance slipped late in training.

With Dhruv Saini, Rohan developed ThinkSwitch to test how much a model that answers directly can learn from one that works through a problem in writing. They trained the direct-answer model on the reasoning model’s final answers after removing its intermediate steps, then blended the updated weights back into the reasoning model. On 30 questions from the 2026 American Invitational Mathematics Examination, the direct-answer model improved from 10 correct answers to 20 while still responding without a written derivation. The separate reasoning model improved from 14 to 22.