Rishith
Prathi

Aug. 2026 – Present

Pittsburgh, PA

Catalyst Lab @ CMU

Machine Learning Research Assistant

• Extending PithTrain, a distributed Mixture-of-Experts (MoE) training framework, with RL post-training for 100B+ parameter models.
• Built a streaming checkpoint loader that loads Hugging Face model weights directly into a fully sharded data parallel (FSDP) model, eliminating a costly offline conversion step and reducing model startup time from 15–45 to 2–5 minutes.

May 2026 – Aug. 2026

San Jose, CA

Adobe

Software Engineering Intern

• Shipped a production RAG context engine with containerized FastAPI serving 1,000+ users, generating AI recommendations for 20k marketing campaigns via Azure OpenAI embeddings and Cosmos DB vector search.
• Created a Scala service, scheduled through Jenkins, that leverages an LLM to cluster AI-agent conversations from Langfuse into weekly Slack reports, saving 10+ engineers a combined 20+ hours/week of manual review.
• Migrated database syncing to distributed Temporal workflows with checkpointing and automatic retries, making syncs resumable after failure and eliminating 5 minutes of redundant work per API deploy.
• Built an evaluation harness comparing 12+ campaign embedding strategies with LLM-as-judge scoring; the top strategy improved average recommendation accuracy by 27% and surveyed user satisfaction by 8%.

Jan. 2026 – Jul. 2026

Pittsburgh, PA

NeuLab @ Language Technologies Institute, CMU

Machine Learning Research Assistant

• Co-designed Forger (2nd author, COLM 2026), an open-source framework that provides reusable workflows — web data retrieval + LLM orchestration — for generating LLM training data at scale.
• Engineered a thread-safe caching layer sharing GPU-hosted, tensor-parallel LLMs across concurrent evaluation workers, eliminating redundant weight reloads and cutting eval runtime 3x.
• Demonstrated Forger-generated data drives a 24-point in-context-learning gain on the HumanEval benchmark through batched vLLM inference with seeded 4-shot subset sampling and bootstrapped confidence intervals.

Jul. 2025 – Dec. 2025

Pittsburgh, PA

Moss Robotics

Software Engineering Intern · Seed-Stage Startup

• Programmed a point cloud training data generator that improved the company's primary ML model accuracy by 4%.
• Boosted YOLO-based neural network performance by 23% by building a C++/Python pipeline that generated and rasterized 5,000+ synthetic images for object detection.
• Validated synthetic data at 89% of real-data accuracy, saving 100+ hours of costly real-world data collection.