Rachel Yang

Ambulance Ride-Along Simulation app preview

Launch the app

See an end-to-end call run

325emergency scenarios vetted against protocols

649textbook / SOP / SOG references

553pages in the medical library

2local EMS Corps piloting

Why I built it

You can’t practice judgment alone

As a first responder and probationary EMT at EVAC, I learned the most by watching experienced EMTs decide which question to ask next, which detail mattered, and what had to happen before transport. Those decisions come from repetition — but in a suburban volunteer corps, a six-hour shift may bring several calls or none.

You can practice using equipment by yourself, but you cannot really practice judgment by yourself. You need a patient who answers back.

So I built one. During a quiet shift, an EMT can open a phone and hear: “You arrive at a home for an 82-year-old woman complaining of weakness. Her daughter says she has not been herself all morning.” What do you do next?

How it works

Run the call

Dispatch to hand-off

Speak or type naturally through scene size-up, patient assessment, vital signs, interventions and transport. The patient, family, dispatcher and scene respond to every decision — what you notice, or miss, changes what happens next.

Get a debrief you can trust

Graded by code, explained by AI

Every call is scored deterministically against an NREMT skill-sheet–style rubric first. Only then does the AI turn the result into a readable debrief — what went well, what was missed, and why. It can improve the wording, never the score.

Choose what to practice

Or let the day choose

Pick from 31 topics and 325 EMT-approved scenarios, or take a random call. Live local weather, season and air quality shift the odds — heavy wildfire smoke makes respiratory calls more likely — and each corps can tune the mix to its own community.

Safety by design

AI that can’t make up medicine

An LLM Wiki underneath. I turned my NYS EMT course notes, protocols and SOPs, chapter by chapter, into a cross-linked knowledge base of 553 pages. Every clinical claim must cite its source; anything the sources can’t support comes back as “not in provided protocols” instead of a guess.
Humans stay in control. The AI drafts new scenarios, but nothing reaches trainees until a certified EMT approves it. Every rejection is distilled into a persistent lesson that guides future drafts, so expert corrections make the system smarter.
Ground truth is data, not model output. Retrieval, grading and citations are checked by ordinary deterministic code, not left to the LLM’s judgment.

Inside the knowledge base · Obsidian

The library is a network, not a list

The whole knowledge base exports to an Obsidian vault. Every call is wired to the topics it teaches and the conditions, procedures and medications it depends on — so thin spots in coverage become visible at a glance.

Force-directed graph of the library: condition pages, practice calls and procedure pages drawn around topic hubs — 553 notes drawn together around 31 topic hubs

Polar chart of the library by type: 368 scenarios, 117 conditions, 55 procedures, 13 medications — The library by type

Radar chart of vetted calls across 31 topics — Topic coverage across 31 topics

Donut chart of vetted calls by kind, dominated by patient care — What each case is built to teach

Bar chart of vetted calls per textbook chapter, mostly nine to eleven each — Vetted calls per chapter — 36 chapters, in the order they are taught

Open source

EMS-AI-SIM

I open-sourced the core infrastructure so any EMS agency can build on it:

Load your own manuals, protocols and SOPs; scenarios cite them page by page
AI-drafted calls sit in quarantine until a certified EMT signs off
One command exports the knowledge base to a connected Obsidian vault
Play in the browser or the terminal — React + TypeScript front end, FastAPI back end

View repository

How I built it

My design, AI-assisted code

The problem definition, architecture, data structures, module design, test plans and clinical review were mine. I used Claude Code to generate the implementation from my specifications, then personally tested on laptop and iPhone and reviewed every change — about 80 pull requests and 160 commits — before merging.

Status · Oct. 2026

In the field, and what’s next

Piloting at two local EMS Corps
Submitted to the Congressional App Challenge
Version 2.0: agency-specific protocols, adaptive practice that targets weak areas, an instructor view, patients whose vitals respond to treatment, and open-source models to cut costs for volunteer corps