Yoga Pose Detection and Correction
Pose detection using feedforward neural networks and Google MoveNet Thunder — 96.5% accuracy across 12 yoga poses, with real-time correction feedback.
READ PAPER ↗TECHNICAL SPECIFICATION · HUMAN, GENERALLY CAPABLE
Mihir
Chauhan
CLASS
ML / AI Engineer — production systems, end to end
TRAINING
M.S. Data Science, CU Boulder · GPA 3.96 · May 2026
MODALITIES
LLM evaluation · Multi-agent systems · RAG · MLOps · AWS
DEPLOYED AT
CU Anschutz · Honda 99P Labs · Kobeyo
LOCATION
Denver metro, Colorado
STATUS
Open to ML / AI engineering roles
FIG. 1 — THE SYSTEM IN QUESTION
0
GPA · M.S. DATA SCIENCE
0%
CV MODEL ACCURACY · PUBLISHED
−0%
QUERY TIME · PRODUCTION PIPELINE
+0%
BUSINESS DATA COVERAGE
Overview
I build AI systems that survive contact with production — from rubric-based LLM evaluation pipelines in clinical workflows to multi-agent platforms running on AWS ECS Fargate.
Right now I'm a research intern at CU Anschutz (LARK Lab), designing evaluation pipelines and guardrails for LLM outputs in clinical annotation, and building LLM-driven infographic generation pipelines. Before that I shipped a multi-agent debate engine for Honda 99P Labs and a self-improving BERT MLOps pipeline at Kobeyo.
Off the clock: football (pretending I'm prime Messi), Colorado trails, and vibe coding at 2 AM with lo-fi on. My specialty is taking a messy, complex idea, turning it into a clean scalable system — and actually finishing it.
Experience
Most recent first. Status tags are accurate.
Projects
Selected systems — with proof where it exists, and none invented where it doesn't.
MORE WORK
Skills
Publications
Building a Real-Time Multi-Agent AI Debate Platform
Technical write-up on architecting a real-time multi-agent debate system with LLMs: agent orchestration, streaming responses, and audience interaction.
READ BLOG ↗Contact
Recruiting, collaborating, or just talking systems — my inbox is open.