I study small language models, and especially what they can learn when data, memory and compute are scarce. I'm a research intern at Stanford's Shah Lab (Stanford Medicine) and Lemons Lab (Graduate School of Education), and a student at Mountain View High School (Class of 2027).

Publications · Website · CV · LinkedIn

All accepted. Abstracts, figures and PDFs are on research.yash-maheshwari.com.

• Halved CLM Exposure Mitigates Late-Training Degradation in Small Recurrent Language Models
  Yash Maheshwari. BabyLM Workshop at EMNLP 2026 (archival). Paper page · PDF · OpenReview
• Below One Bit: Training Route and Budget Shape Robustness Under On-Device Storage Constraints
  Yash Maheshwari. NeurIPS 2026 Workshop on On-Device Intelligence (ODI), poster. Paper page · OpenReview
• Training and comparison of fine-tuned small language models as math tutoring assistants
  Yash Maheshwari, Rinky Gupta. Journal of Emerging Investigators, in press. Paper page · Code

• Shah Lab, Stanford Medicine (with Prof. Nigam Shah). I maintain HealthAdminBench and am building its second version, a benchmark of AI agents doing healthcare administrative work such as prior authorizations and denial appeals inside a realistic Epic EHR.
• Lemons Lab, Stanford Graduate School of Education (with Prof. Chris Lemons). I built the evaluation suite for Kai, an AI reading tutor used by 1,200+ students in 10+ districts, and cut its response latency by 75%. I also work on PAWS, a handwriting tutor for kindergarteners on iPad.
• Independent research on pretraining under tight data budgets and sub-1-bit compression. I train these models from scratch on a MacBook with MLX.

• Patents (filed 2025): Hierarchical Aggregation Tree for MCP Server Selection and Execution (non-provisional, sole inventor) and Predictive Compliance for AI Agents (provisional, co-inventor with Aisera).
• Industry: AI Engineering Intern at Aisera (2025), where I built MCP servers for Salesforce, Clari and Slack and an open-source MCP bridge, and placed 2nd of 30 teams in the company hackathon. AI Research Contractor at Kinetic Systems (2026).
• Speaking: an upcoming keynote at FETC 2027 (January), two keynote panels at the Common Sense Media Summit 2026, and talks at FETC 2026, the ASU+GSV Summit 2025 and Google.
• Press: featured in The Washington Post (Oct 2025) for helping write my school district's AI policy.
• Community: co-founded Tech Spark, a 501(c)(3) that has run five summers of K-8 robotics and coding; software lead for FRC 9584, which won the Highest Rookie Team Award in its division at the 2024 World Championship.

Tools I use most: Python, MLX, Hugging Face (Transformers, TRL), TypeScript and React.