Research Scientist & ML Engineer
Rohan Pandey
ML, mathematics & computational biology — and founder of two startups spanning rocket recovery and AI-powered research.
The growing map — hover any point to read it, drag the timeline to replay it.
Now
- Founding Qorbit Systems & Lumina Research
- Building at DigitalOcean — dedicated GPU inference (vLLM, prefill/decode disaggregation)
- Research in RL post-training & verification — NeurIPS 2026 workshop paper on reward hacking in RLVR
- Speaking — AI Engineer Code Summit talk on LLM settings that only win on paper
- Mentoring at Anchor High — research, internships & admissions
Ventures
Two companies. One orbit — hardware that saves rockets, and software that ships papers.
Qorbit Systems
Smarter recovery for model rocketry — guided pre-flight checks, sub-400ms backup deployment, and shareable flight records for TARC & competition teams.
- 4,000+ competition teams
- <400 ms recovery response
- $240M addressable market
Lumina Research
A project-based course that takes you from a research question to a submitted paper — AI-powered, structured, and built for people who want to actually publish.
- End-to-end research workflow
- AI-guided at every stage
- Question → paper in one arc
About
I'm an M.S. student in Applied & Computational Mathematical Sciences (ACMS) at the University of Washington, where I also completed my B.S. in 2026, and a Software Engineer on DigitalOcean's Dedicated Inference team. I minored in Neural Computation and Engineering, with a focus on AI in neuroscience.
I'm also the founder of Qorbit Systems (smarter recovery for model rocketry) and Lumina Research (AI-powered research education). My technical work spans RL post-training and reward hacking in RLVR, reinforcement learning for polynomial synthesis (GNNs, PPO, MCTS), GPU inference serving, and cancer modeling for CAR T-cell therapy.
When I'm not in the lab or building rockets, I'm on the tennis court or winning at Monopoly. I'm always up for collaborating and exploring new tech.
Skills & expertise
Experience
Software Engineer II · DigitalOcean Mar 2026 – Present
Dedicated Inference team: benchmarking, serving architecture, and performance for DigitalOcean's dedicated GPU fleet.
- Own inference benchmarking: orchestrate vLLM on Kubernetes to measure throughput and latency across hardware, model, and concurrency configurations; results set capacity, routing, and SLA decisions
- Built DigitalOcean's first prefill/decode (PD) disaggregation setup with KV cache reuse via LMCache, cutting time-to-first-token and raising sustained throughput
- Building a Model Optimization Tool that profiles and tunes GPU kernels via LLM tool-use, automated benchmarking, and structured search
- Own fleet observability (Prometheus, Grafana, PagerDuty) and debug GPU, Linux, and network performance regressions
- Co-presented Stop Shipping LLM Settings That Only Win on Paper at AI Engineer Code Summit
Founder · Lumina Research Mar 2026 – Present
Project-based course teaching AI-powered research — from formulating a question to submitting a paper.
- Structured workflow covering literature review, analysis, and writing
- Built for researchers who want to ship real publications
Founder · Qorbit Systems Dec 2025 – Present
Building smarter recovery for model rocketry — guided pre-flight workflows, sub-400ms backup deployment, and shareable flight records for TARC & competition teams.
- Physics-engine flight simulation and real-time telemetry platform
- Designed for 4,000+ competition teams in a $240M global market
Lead Reinforcement Learning Researcher · Math AI Lab, UW Sep 2025 – Present
Formulated arithmetic-circuit synthesis as a single-player RL game; a PPO + Graph Neural Network agent reaches ~70% success on degree-3 polynomials via curriculum learning, shaped rewards, and symbolic verification.
- Scaled to higher-degree synthesis with MCTS + AlphaZero-style self-play and a reusable library of factorizable subexpressions as recursive subgoals; ablated against SAC baselines
- First author at ICML 2026 (AI for Math Workshop) and ICLR 2026 (RSI Workshop)
ML Engineer · Mercor Dec 2025 – Mar 2026
Built and deployed an end-to-end ML pipeline for large-scale text sentiment analysis, enhancing prediction accuracy through strategic model selection.
- Fine-tuned and evaluated transformer models on domain-specific data using custom scoring metrics
- Owned data preprocessing, model validation, and inference automation for production-ready workflows
Cloud & Digital SAP Intern · PwC Jun – Aug 2025
Engineered high-performance Python data pipelines leveraging vectorization and parallel I/O to reduce latency; built a secure self-service web app for automated workflow configuration and monitoring.
Applied Mathematics Researcher · Fred Hutch Aug 2024 – Dec 2025
Developed mathematical models of CAR T-cell and tumor interactions using systems of ODEs trained on patient-derived B-ALL data, applying regression and loss optimization to evaluate treatment efficacy.
ML Intern · Naval Surface Warfare Center Oct – Dec 2024
Built a reinforcement learning framework with DDPG agents in MATLAB and Simulink, optimizing aerodynamic performance on axial turbomachinery simulations under complex physical dynamics.
ML Intern · MINDCO LABS Sep – Dec 2024
Engineered a real-time eye-tracking system integrating EEG headset signals with deep learning to decode gaze to text, contributing to brain-computer interface (BCI) research.
Project Manager Intern · NASA L'SPACE May – Aug 2024
Lucy & Artemis; led team of 11 on lunar ice mapping rover; PDR, budget, programmatics.
Lead Engineer · SARP Oct 2022 – Jun 2024
ARES: GPS-guided recovery, drogue logic, telemetry in C++/Python.
Hackathons
Builds with real-world impact.
Husky Maps
Faster way to commute through campus — shortcuts through buildings for less walking time.
Winner — UW Claude Hackathon 2025BloomSync AI
AI-guided bioinformatics pipeline that reprograms apple flowering genes through FT/TFL1 edits — new-wood fruiting and robotics-ready orchards.
2nd Place — DubHacksNeuralTrace
Real-time EEG learning visualizer — live EEG from OpenBCI in the browser via Web Serial API, 3D brain with electrode glow, and interactive neural connection animations. No backend required.
Top project — NeuroAI HackathonPublications
Papers, preprints, and research posters — all in one place.
View all research
17 papers and preprints, including first-author workshop papers at NeurIPS, ICML, and ICLR 2026, plus research posters.
- NeurIPS 2026
- ICML 2026
- ICLR 2026 ×2
Mission Control
Pilot the rocket through the asteroid fields.
Contact
Collaborate or say hi.