SALT SHAKER STUDIO

It's been a couple months. Not because nothing happened — but because im lazy with writing posts. Here's the update:

1. ABG Beta: Feature Complete
All four modes are built: Foundational, Beginner, Stepwise, and Expert. The logic is solid, the cases are solid, and the beta build is done. The frontend UI is totally refreshed and fancy (thanks Fable, RIP). I apparently have to wait for some ID number to submit to the app store (??), so thats what I'm doing. Also waiting for a pulm/crit attending review of the app (looking at you, Harman). C'mon bro.

2. About the Version on the Site
You're probably thinking: wtf, the ABG app on this site is the exact same.... That is true, the newer shiny app with the glorious and beautiful UI will be available on the app store. Everything functions the same otherwise, so if you don't mind risking some eyeball cancer, give it a shot.

3. Also, the Website Looks Different Now
If you haven't been here in a while, yes, it's intentional. We went from full Windows 95 (never go full windows) to full Windows 98 update. We all know Linux is better but this is nostalgic. (Protip: try the start menu at the bottom left).

If you've been using ABG for a while and the step-by-step guidance has started to feel like training wheels, this update is for you.

What Expert Mode Is
Expert Mode strips everything back. You get a blood gas. No steps. No hints. No hand-holding. Just the numbers and a text field. Interpret it correctly and move on. Get it wrong and you'll find out. It's designed for people who already understand the logic and want to test themselves against real-world-style cases without the scaffolding.

Why We Added It
The core ABG app was always built around teaching the process — walking you through pH, PaCO2, HCO3, compensation, one step at a time. That's still there and it still works. But we kept hearing from more advanced users that they wanted something closer to what it actually feels like on the wards: ABG in hand, attending waiting, no prompts. Expert Mode is that.

Give it a try and let us know how it goes. The bar is high — no pun intended.

Quick background: I've been building teaching tools for medical trainees for a while. At some point I got curious — what if instead of hard-coding decision trees, I just trained a small language model to reason through clinical cases the way a physician would? Reader, I tried it.

1. What I Actually Did
Took Llama 3.2 3B Instruct — a small, open-source model — and fine-tuned it on 511 physician-annotated cases from the DiReCT dataset using LoRA. Ran it on a Google Colab T4 GPU for 3 training epochs. The loss dropped from 22 to 3.7, which sounds like a lot because it is. I did not invent any of the underlying technology here. I just pointed existing tools at a medical dataset and let it cook.

2. Does It Actually Work?
Yes, actually. I gave both the base model and my fine-tuned version the same clinical vignette — a classic hypothyroidism presentation. The base model confidently stated that cold intolerance is caused by too much thyroid hormone. Which is the opposite of true. It also said the patient had weight loss. Hypothyroidism causes weight gain. These are not subtle errors. These are the kind of errors that would make an attending physically wince.
The fine-tuned model correctly identified hypothyroidism, flagged Hashimoto's as a likely cause, and walked through a logical flow: symptoms → testing → treatment. Basic? Sure. But correct. That's not nothing.

3. This Is a Proof of Concept, Not a Product
To be very clear: this is early and experimental. Training data is limited to 511 cases, the base model is 3B parameters, and there's a long road ahead. What I wanted to know was whether this direction was even worth pursuing. The answer seems to be yes. Next steps include expanding the dataset, upgrading the base model, and cleaning up the hedging language the outputs love to pepper in. Nobody needs a model that qualifies every sentence with "however, it is important to note that I am an AI."

Model is live on HuggingFace if you want to poke at it: huggingface.co/ploppy2/Clinical-Reasoning-Test1

Three things happened. Not all of them are huge. All of them are worth knowing about.

1. SAL: A New App for Learning About Hyponatremia
I decided to create SAL, working title, don't @ me. It's a teaching tool for hyponatremia, one of the most annoying electrolyte disorders in medicine. Study it 100 times and still get confused. Good thing this app generates infinite questions. The app is in early development, but the core concept is the same as ABG: walk you through the logic step by step until it clicks. More updates soon.

2. ABX Alpha Prototype Is a Thing That Exists Now
We've put together an alpha prototype for ABX — the antibiotic decision-making tool — and this one is built in Godot. It is rough. It is very much a prototype. The entire point right now is to get something in front of people and see if the core concept holds up before we build the real thing around it.

3. Logos Got Updated
The SSS and ABX logos have been updated. They look better now. I forgot to add "SAL" to the SAL logo... oops.

The ABG prototype is officially deployed to saltshakerstudio.net/abg. No app store. No "enable unknown sources" deep in your Android settings. Just a link. Click it. It works on your phone. We're living in the future, bro.

1. What's Actually There Right Now
The prototype has the core ABG interpreter engine, the random case generator, and the foundational and beginner teaching modes. It is not finished. It is not polished. There are almost certainly bugs. That is the point. If you wait until something is perfect to share it, you will be sharing it never.

2. Why Web Instead of a Native App
Short answer: because deploying a Flutter web build to GitHub Pages takes one terminal command and zero App Store review cycles. Long answer: same as the short answer. Native builds are coming, but getting the core learning experience in front of real residents now is more useful than spending two weeks fighting Xcode provisioning profiles.

3. What I Actually Want From You
Use it. Break it. Tell me what feels wrong. Specifically: does the stepwise logic make sense the way a real attending would walk you through it? Are the cases realistic? Is there a disorder pattern you keep seeing on rounds that the generator never seems to produce? Yell at my on twitter or email - links at the bottom of the page.

4. What's Next
Planning to add a pro mode for you giga-chads out there that dont need no tutoring so you can practice solving ABGs without any help. At some point will also add a leaderboard so you can be a gunner secretly. Don't pretend you wouldn't want to flex on the OTHER medical school in your area.

Go break something: saltshakerstudio.net/abg

I've just pushed a massive update to the ABG Tutor. My goal was simple: make every case feel less like a guessing game and more like a clinical reasoning script—basically, a digital version of your attending staring at you until you remember Winter's formula.

1. Hand-Holding for Simple Disorders
I've built dedicated, stepwise flows for all seven "simple" patterns. Instead of just hitting "reveal" and seeing a diagnosis you don't understand, the Tutor now walks you through a 5-step scaffold. It'll nudge you through pH, primary processes, and the "acute vs. chronic" bicarbonate hint until you reach the right answer. It's like training wheels, but for people who carry pagers.

2. Mixed Disorders (The "Where Am I?" Phase)
Mixed metabolic and respiratory issues are usually where the panic sets in. I've added structured 6-step paths that force you to actually use the formulas. The Catch: For things like Winter's formula, I'm giving you numeric multiple-choice options without showing the equation. You actually have to know it. (I know, I'm a monster.)

3. Triple Disorders: Now With Actual Logic
The triple-disorder engine used to be a bit... "random-ish." You might have seen an explanation for HAGMA + Met Alk while the numbers were clearly screaming NAGMA. I've rewritten the logic from the ground up. The generator now picks a logical pattern first, then forces the math to obey the laws of physics. Whether it's Delta-Gaps or expected PCO₂, the numbers will actually make sense now.

4. The New 7-Step Triple-Disorder Flow
Triple cases are now a full 7-step journey through: anion gap status, the Delta-Gap interpretation (the "hidden" stuff), Winter's formula (the math stuff), and the final, glorious, multi-hyphenated diagnosis. Everything is computed dynamically, so the explanation will finally match the numbers on your screen. No more gaslighting from the software.

Welcome to the new Salt Shaker Studio website! We've gone full retro with a modern functional design. This site will be our home for development updates, game announcements, and behind-the-scenes looks at how we build medical education games.

Stay tuned for regular dev updates as we continue building ABG, ABX, and GAS.

We've completed the core blood gas parser engine and the random ABG generator engine; as well as foundational and beginner modes. We are now building out the stepwise teaching mode. After that the plan is to move forward with either a more advanced teaching mode and then expert mode.

Should I add a leaderboard so you can mog your classmates? Let me know!

Arman here. I'm thrilled to officially launch Salt Shaker Studio. As a medical educator, I've seen firsthand how traditional medical education can fall short. Lets be real, nobody likes to learn by being read a PowerPoint presentation by someone who didn't even make it! My mission is to create engaging, interactive games and apps that make learning medicine genuinely fun and actually effective so you don't waste your precious time. Theres a reason it's easy to remember the names of 151 Pokemon but hard to remember the functions of 151 antibiotics!!