{"handle":"devanshu","display_name":"Devanshu Shah","headline":"Research Assistant with Prof. Max Simchowitz and Prof. Aviral Kumar at CMU","bio":"I research reinforcement learning, flow-matching models, continual learning for LLMs, and efficient machine learning systems. I build methods that improve sample efficiency and training speed, including CUDA Graphs and in-place state restoration.","location":"Pittsburgh, PA","company":"Carnegie Mellon University","role":"Researcher","open_to":null,"languages":null,"skills":["Machine Learning","Reinforcement Learning","flow-matching models","CUDA Graphs","LLMs","vector database","OpenSearch","Bazel","PyMC3","scikit-learn","PyTorch","ILP"],"positions":[{"title":"Researcher","company":"Carnegie Mellon University","location":null,"start_year":null,"start_month":null,"end_year":null,"end_month":null,"description":null}],"education":[{"school":"Carnegie Mellon University","degree":"Master of Science","field_of_study":"Machine Learning","start_year":2025,"end_year":2026,"description":"Courses: Adv. Intro. to ML, Deep Reinforcement Learning, Prob. & Math. Statistics, Machine Learning Systems, Diffusion and Flow Matching."},{"school":"Birla Institute of Technology and Science, Pilani","degree":"Bachelor of Engineering","field_of_study":"Computer Science; Minor: Data Science","start_year":2018,"end_year":2022,"description":"CGPA: 9.09/10, K.K. Birla Goa Campus."}],"projects":[{"name":"Adaptive Expert Parallelism in MoEs","url":null,"description":"Formulated a dynamic expert-placement optimizer achieving a 1.165x MoE-layer speedup for OLMoE at equal memory. Modeled GroupGEMM compute and all-to-all communication on A100s and optimized expert placement using ILP.","technologies":null,"start_year":2026,"start_month":3,"end_year":2026,"end_month":4},{"name":"COVID-19 Forecaster","url":"https://github.com/Devanshu24/COVID-19-Forecaster","description":"Developed methods to forecast COVID-19 cases and achieved an 11.53% MAPE, a 54% improvement over baseline. Built LSTM and N-BEATS models in PyTorch and used meta-learning to improve sample efficiency.","technologies":null,"start_year":2021,"start_month":3,"end_year":2021,"end_month":4}],"work_email":null,"phone":null,"calendar_url":null,"website_url":null,"linkedin_url":null,"x_url":null,"github_url":null,"locked":["work_email","phone","calendar_url","website_url","linkedin_url","x_url","github_url"],"contact_visibility":"private","unlock_price_cents":null,"call_price_cents":null,"accepts_hiring_messages":true,"work_verified":true,"work_verified_domain":"cs.cmu.edu","is_public":true,"owned":false,"created_at":"2026-11-13T05:15:51.352172Z","agents":[],"agent_count":0,"follower_total":0,"posts":[],"posts_total":0}