Rohan Sinha

CS @ Cornell

Machine learning, AI systems, computational research.

GitHub LinkedIn

Experience

Cornell University

Reinforcement Learning Engineer

Sep 2026 – PresentIthaca, New York, United States

Developed reinforcement learning controllers for 2D fluidic pinball using Python, PPO, and HydroGym, reducing simulated drag by 35–42% versus uncontrolled flow on held-out test states.
Built PPO and gradient-enhanced PPO comparisons across four Kolmogorov flow configurations using JAX and Flax, reducing benchmarked GPPO update time by 27% through optimized gradient checkpointing.
Engineered HPC training pipelines across four NERSC A100 GPUs using Slurm, with resumable model and optimizer checkpoints, finite-difference gradient verification, and automated simulation benchmarks.

Temple Allen Industries

Machine Learning Engineer

Feb 2026 - Aug 2026Rockville, MD

Architected reinforcement learning control pipeline in Python and OpenAI Gym to minimize actuator motion variance under external disturbances, reducing positional discrepancy by 38%.
Evaluated and fine-tuned computer vision detection models for aircraft component recognition using PyTorch and OpenCV, improving classification accuracy by 14% across validation datasets.
Developed simulation and inference testing framework with NumPy and real-time sensor feedback analysis to benchmark actuator stability and vision model robustness under dynamic operating conditions.

University of Maryland

Computational Number Theory Researcher

Apr 2025 - May 2026Remote

Co-author of 2 accepted ACM SIGACT News open problems and arXiv paper on congruence monoid and Gaussian primes.
Built custom sieving algorithms and modular filters achieving 3% MAPE and derived 2 prime density formulas validated via nonlinear least-squares regression.
Optimized Miller-Rabin tests with Numba, cutting per-check latency ~40% on 10⁹+ candidates and verified densities using a Python/NumPy analysis framework.

University of Maryland

Computational Game Theory Researcher

Mar 2024 - Feb 2026Hybrid

First author of University of Maryland-accepted paper analyzing player skill and ideal positions in imperfect Nim through 10⁶+ Monte Carlo simulations in Java.
Designed 4 strategy models with up to 100% accuracy using algorithms for move optimization, win detection, and stochastic modeling.
Applied NumPy, SciPy, and Matplotlib for data analysis; leveraged virtual CPUs, parallel processing, and Gaussian smoothing to optimize runtime by ~26% and process 10⁹+ 3D points.

The George Washington University

NLP Researcher

Nov 2024 - Jun 2026Washington, DC

First author of manuscript quantifying semantic drift during COVID-19 via 6-trait NLP pipeline on 22.2k Reddit Q&A pairs (2019-2023).
Implemented GPT-4o few-shot classifier generating 133.3k Likert scores, achieving 0.82 Krippendorff's alpha and 15% higher calibration accuracy using isotonic regression.
Applied Bayesian change-point detection, PELT, and causal impact forecasting to identify post-2020 linguistic shifts with NumPy and scikit-learn.

Education

Cornell University

Computer Science · 4.0 GPA

Expected graduation: 2029

Research

Working in Cornell’s Co MSAIL Lab on using reinforcement learning agents for fluid dynamics.

Coursework

OOP I, OOP II, Calculus I, Calculus II, Data Structures & Algorithms

Projects

Adaptive gRPC Compression

Reduced p95 RPC latency 61.7% across 161K calls by training LinUCB to select gzip/zstd from runtime context.
Saved 863.7M payload bytes (56.7%) across 161K RPCs by adapting compression to entropy and CPU load.
Achieved 1.38 µs decision latency by embedding bounded LinUCB inference in Go and streaming gRPC interceptors.

Go
gRPC
LinUCB
Contextual Bandits

View on GitHub

ML CI Test Conflict Detection Scheduler

Instrumented 280 pytest cases for CI conflicts by tracing files, sockets, ports, DBs, and processes with Rust eBPF.
Scheduled CI across 8 workers with 3 risk policies using conflict probability, overlap duration, and LPT placement.
Built a Rust eBPF/Python/Go/Postgres pipeline to trace 280 tests, predict conflicts, and schedule workers.

Rust
eBPF
Python
Go
Postgres

View on GitHub

GroundSense

Architected React/TypeScript + Supabase/Postgres platform with RLS, Edge Functions, and a source-audit ledger.
Built Deno ETL normalizing 2K+ BLS/EIA/FRED/UN records with Gemini-assisted claim extraction pipelines.
Mapped external shocks to calibrated exposure scores and mitigation measures across 5+ supply-chain workflows.

React
TypeScript
Supabase
Deno
Gemini

View on GitHub

Federated Learning for Clinical NLP

Built PyTorch + Flower (FedAvg) simulation across hospital nodes over 50-100 rounds; applied Opacus differential privacy (ε 1.5-3.0) and secured via TLS 1.3, AES-256, and JWT auth.
Trained ClinicalBERT/mBERT (~110M params); reached 80%+ F1 on holdout with under 15% communication overhead vs centralized baseline, robust to client dropout and label noise.

PyTorch
Flower
Opacus
ClinicalBERT

View on GitHub

More projectsFewer projects

Multi-Agent RL for Energy Distribution

Developed multi-agent PPO system with a custom 2-3 home + 1 battery environment and peak/off-peak pricing; ran ~1e6 env steps with 8-16 rollout workers and AMP fp16.
Delivered 15-20% cost reduction vs baseline across 100 eval episodes with under 1% reward variance; resilient under 2x demand surges at ~25k env-steps/s throughput.

Ray RLlib
PyTorch
Gymnasium

View on GitHub

Cross-Lingual Medical NLP

Built a clinical NLP transfer framework using Hugging Face Transformers, PyTorch, LoRA, and Gradio.
Fine-tuned mBERT on 10k+ English medical NER samples and adapted it to Spanish with 200 labeled examples for low-resource transfer.

Transformers
LoRA
Gradio

View on GitHub

All repositories

Publications

ACM SIGACT News, 2026

Estimating the Number of Primes in Unusual Domains

Investigates the distribution of primes beyond the natural numbers: congruence monoids, the Gaussian integers, and quadratic extensions of Z, with heuristic conjectures backed by empirical density evidence.

University of Maryland

Comparative Analysis of Ability and Ideal Starting Positions in 1 and 2 Pile Nim Games

Monte Carlo simulations and a derived mathematical model showing that under imperfect play, player ability outweighs starting position as pile size grows. Accepted after formal peer review.