Inspiration
CPR is a practical skill, but learning it effectively requires more than being told whether an attempt was good or bad. A trainee may receive feedback such as "your compression rate is too fast" or "improve your hand position," but that still leaves an important question unanswered:
What should I practice next, and did that practice actually help me improve?
This inspired ResusLoop.
Rather than building another system that simply analyzes CPR and produces a score, ResusLoop is designed around a closed learning loop:
Perform → Observe → Diagnose → Explain → Practice → Reassess → Improve
The goal is to explore how AI can move beyond passive assessment and become an adaptive training assistant for physical skills.
What it does
ResusLoop is an evidence-grounded AI CPR practice coach.
A trainee records a CPR practice attempt using a camera. The system analyzes measurable aspects of the performance such as:
- compression rate and rhythm,
- hand-placement region,
- upper-body and arm posture,
- movement consistency.
Each observation is also associated with a confidence level.
This is important because ResusLoop is deliberately designed not to pretend that every physical measurement can be reliably inferred from ordinary video. For example, if a measurement such as physical compression depth cannot be determined reliably from the available camera view, the system should mark it as unavailable or uncertain rather than presenting a misleading precise value.
The reliable observations are then combined with authoritative CPR guidance and passed to an NVIDIA-powered reasoning agent.
Instead of overwhelming the trainee with every detected problem, the agent identifies the highest-priority reliable weakness and selects an appropriate corrective practice drill.
The trainee completes the drill and performs another CPR attempt.
ResusLoop then compares the measurable performance between the attempts and updates the trainee's skill profile.
The result is not simply:
CPR → Score
but:
Observe → Measure → Reason → Retrieve Evidence → Practice → Reassess → Learn
How we are building it
ResusLoop separates perception, reasoning, evidence, and training decisions instead of asking a single AI model to perform the entire task.
1. CPR Perception Layer
Computer vision analyzes the CPR recording and converts movement into structured observations such as compression timing, hand location, posture, and movement consistency.
2. Confidence-Aware Measurement
Every measurement is evaluated for reliability before it can influence coaching.
Low-confidence observations are prevented from becoming strong conclusions.
3. Physical Video Reasoning
NVIDIA Cosmos is being explored for higher-level reasoning over selected video segments where contextual understanding of physical movement can complement deterministic computer-vision measurements.
4. Evidence Retrieval
When a technique problem is detected, ResusLoop retrieves relevant information from trusted resuscitation guidance rather than allowing the language model to invent CPR recommendations from memory.
5. CPR Training Agent
NVIDIA Nemotron, running through Nebius infrastructure, acts as the reasoning layer.
It receives:
- reliable performance measurements,
- measurement confidence,
- relevant video observations,
- retrieved evidence,
- and previous training history.
It then determines which skill should be addressed next and selects an appropriate corrective drill.
6. Closed-Loop Reassessment
After the drill, the trainee performs another attempt.
Objective measurements — rather than the language model's opinion — are used to determine how performance changed.
The result is stored in the trainee's skill history so that future sessions can focus on the next meaningful weakness.
Why NVIDIA and Nebius matter
NVIDIA and Nebius are not being added simply to satisfy a technology requirement.
NVIDIA Nemotron provides the agentic reasoning required to interpret structured observations, consider evidence, prioritize weaknesses, and determine the next training action.
NVIDIA Cosmos is being investigated as a multimodal physical-reasoning component for selected CPR video segments.
Nebius Token Factory / AI Cloud provides the inference and AI infrastructure used to run these NVIDIA models and support the agent workflow.
The architecture is intentionally designed so that AI reasoning is separated from objective measurement.
The model can reason and explain.
It cannot declare its own answer correct.
Challenges
One of the central technical challenges is that video-based physical-skill assessment can easily create false precision.
Some characteristics, such as compression rhythm, can be estimated reasonably well from temporal movement. Others — particularly physical compression depth from a single uncalibrated camera — are significantly more difficult.
Instead of hiding this limitation, ResusLoop makes uncertainty part of the system architecture.
Another challenge is preventing an AI coach from producing unsupported medical-training recommendations. ResusLoop therefore separates evidence retrieval from reasoning and is being designed to ground coaching decisions in authoritative CPR guidance.
A third challenge is scope. ResusLoop is being developed as a solo hackathon project, so the first version intentionally focuses on a small number of measurable CPR skills and one complete training loop rather than attempting to automate every aspect of CPR assessment.
What I am learning
Building ResusLoop is an opportunity to explore how several AI techniques can work together as one reliable system rather than as isolated demonstrations.
The project involves learning and applying:
- computer vision for physical movement analysis,
- multimodal AI,
- agentic workflows and tool calling,
- NVIDIA Nemotron and Cosmos models,
- Nebius AI infrastructure,
- evidence retrieval,
- confidence-aware AI design,
- longitudinal skill memory,
- and evaluation of AI-assisted physical training.
One of the most important ideas behind the project is that a useful AI system should know the difference between what it observed, what it inferred, and what it cannot reliably determine.
Responsible Scope
ResusLoop is intended as a formative CPR training aid, not a medical diagnostic system, emergency-response system, or replacement for certified CPR instruction.
The prototype focuses on helping trainees practice measurable aspects of CPR technique and understand what they should work on next.
What's next
The first goal is to demonstrate one complete and reliable loop:
CPR Attempt → Measurement → Evidence-Grounded Diagnosis → Targeted Drill → Second Attempt → Measured Change
Once this core loop is validated, future work could explore additional camera configurations, sensor-assisted measurements, more advanced physical reasoning, and adaptation of the same closed-loop architecture to other physical training skills.
Log in or sign up for Devpost to join the conversation.