Building MedixAI in Public, Week 1: The SOAP Engine That Wants to Kill Paper in Every Filipino Clinic
A product feature report from the desk of the founder. First in a weekly series documenting the build of MedixAI, an AI clinical operations platform made for the Philippines. — https://getmedixai.com
Why I am writing this
I grew up watching my family wait. Wait in line at the barangay health center. Wait for a doctor who had fifty other patients behind us. Wait for lab results printed on thermal paper that faded before the next visit. Wait for a diagnosis that sometimes came too late, sometimes came wrong, and almost always came after the doctor had already spent half the consultation writing notes instead of looking at the person in front of them.
That waiting is not a scheduling problem. It is a data problem. And nobody in Philippine healthcare has treated it like one.
MedixAI — https://getmedixai.com is my attempt to fix it. Not with a chatbot. Not with another booking app. With actual clinical infrastructure that learns every patient, remembers every visit, and gives Filipino doctors back the one thing they never have enough of: time with the patient in front of them.
This is the first of what will be a weekly published report on what we are shipping, what is breaking, and what we are learning. No marketing gloss. No AI hype. Just the build, in public, in real time.
The tagline of the product is “No pressure. Just help.” That is also the rule for this newsletter.
What we shipped this sprint: the SOAP Auto Generation Engine
Every consultation in a Philippine clinic ends the same way. The doctor picks up a pen, opens a manila folder, and writes a SOAP note by hand. Subjective. Objective. Assessment. Plan. Four sections, five to ten minutes, multiplied by thirty patients a day. That is two and a half hours of documentation per doctor, every single day, stolen from patient care and from their own families.
Our Week 1 sprint replaces that entire ritual with one interaction.
The doctor types two sentences. Drops in the x-ray the patient brought from another hospital. Attaches the prescription from a specialist the patient saw last month. Taps generate. Ten seconds later, a complete, editable, PhilHealth compatible SOAP note is on the screen, with a prescription draft, a follow up schedule, and a flag if anything we saw in the uploaded images contradicts what the doctor is about to prescribe.
That is the feature. That is the whole pitch.
The three inputs
We intentionally kept the inputs to exactly what a Filipino doctor already has on their desk at the moment of the consultation.
The patient intake. This already runs live in MedixAI. Before the patient even sees the doctor, they complete a Tagalog and Taglish conversational intake on their phone or a tablet at the reception desk. Chief complaint, symptoms, duration, allergies, current medications, relevant history. By the time the doctor calls them in, a structured clinical brief is waiting.
The doctor’s typed notes. One to five sentences. That is all we ask. “Tonsils inflamed, no fever, amoxicillin 500mg three times daily for seven days, follow up in one week.” Any Filipino GP can type that in fifteen seconds.
The uploaded images. This is the part that makes doctors lean forward when we demo it. Every Filipino clinic has a stack of paper artifacts the patient walked in with. An x-ray from the ER last Tuesday. A lab result from a diagnostic center across the street. A prescription from a specialist in another province. A photo of a rash that looked worse yesterday. The doctor snaps a photo, tags the category, and drops it into the consultation. We do the rest.
What the engine actually does with those inputs
This is where the work lives.
We built the engine on Claude Sonnet 4 with vision, running through Supabase Edge Functions. When the doctor taps generate, we pass a carefully constructed multimodal prompt that fuses the intake, the typed notes, and every uploaded image. Each image is analyzed according to its tagged category. X-rays and clinical photos are described objectively, never diagnosed. Lab results have their values extracted and flagged against normal ranges. Prior prescription photos have their drug names pulled and cross referenced against anything the doctor is about to write.
That last point is the one I lose sleep over. Polypharmacy kills Filipinos every year who see four different doctors none of whom know what the others prescribed. If our engine flags one drug interaction that a tired doctor at 7pm on a Friday would have missed, the entire product has justified its existence.
The output is a full SOAP note, a prescription draft, a follow up record, a list of image findings, and a red flags array the doctor can scan in three seconds. Every section is editable inline. Edits are tracked so we can feed them back into the prompt and improve the engine over time. One tap finalizes the note, saves it to the patient timeline, and exports a PhilHealth formatted PDF with the referenced images as an appendix.
Every image the doctor uploads becomes a permanent part of the patient’s longitudinal record. Searchable by category, by date, by consultation. The next time that patient walks into any clinic running MedixAI, the doctor sees the entire history, not a stack of paper the patient may or may not have remembered to bring.
That is the quiet thing nobody is building in the Philippines. A real longitudinal health record. One patient at a time. One clinic at a time. Starting now.
The time saved counter
There is one feature in this sprint I will not cut under any circumstance, and I want to explain why.
Every time a doctor finalizes a note, we measure the wall clock time from the moment they tapped “Add Consult Notes” to the moment they tapped “Finalize.” We compare it against a twelve minute baseline, which is what manual SOAP documentation plus paper filing actually costs a Filipino GP. The delta gets displayed on their dashboard as a running daily total.
“You saved 87 minutes today.”
That is not a vanity metric. That is the retention engine. That is ninety minutes of dinner with their kids. That is ninety minutes less burnout. That is the number the doctor screenshots and sends to the colleague who has not signed up yet. Every SaaS product that has ever worked in healthcare found a way to make the user feel the value in a visceral, countable, shareable way. This is ours.
The technical bets
A few decisions worth naming because they will matter when the product starts to scale.
We chose a Progressive Web App, not native mobile. Every health app that ever launched in the Philippines bled weeks waiting for Apple’s review queue. A PWA installs from a link, works offline for note editing, updates instantly, and runs on any five inch Android phone a barangay health worker already owns. The day we have enough doctors to justify a native app is the day we build one. Not before.
We chose text and images first, voice second. The original plan had Whisper transcription in Week 1. I cut it. Here is why. Voice transcription in Taglish is still a fragile problem, and every minute we spend debugging it is a minute we are not closing the first five paying doctors. Text input is faster than voice for anyone who types well, and most doctors type fine. Images give us longitudinal memory immediately. Voice is Phase 2, after twenty doctors are paying and the core loop is proven. Discipline over ambition.
We chose Claude Sonnet 4 with vision for the multimodal layer. We recently migrated the platform off GPT-4o to Claude, and the vision capabilities are a big part of why. Conservative clinical language, strong refusal behavior on hallucinated findings, and a single API call that handles text plus multiple images in under twelve seconds at our latency targets. The system prompt is written to be clinically conservative by design. Images are described, never diagnosed. The doctor is always the decision maker. The AI is the assistant that never gets tired.
We chose Supabase as the entire backend. Postgres, Storage, Edge Functions, Auth, Row Level Security, all in one stack, data residency in Singapore for ASEAN compliance. One vendor to reason about. One bill to pay. One dashboard to monitor. When you are a small team building for a large country, you do not buy complexity you cannot afford to maintain.
What is hard about this and what can still go wrong
I said no marketing gloss. So here is the honest list.
Clinical accuracy is a moving target. We have a quality bar for v1: eighty percent of generated SOAPs must require fewer than three edits from the doctor, zero hallucinated findings, one hundred percent red flag detection on known red flag cases. Our synthetic eval suite has twenty test cases covering common Filipino GP scenarios and every image category combination we support. If the engine fails the bar, we do not ship. We iterate.
Image quality is unpredictable. A photo of a lab result taken under a fluorescent tube in a crowded clinic at 4pm is not the same as a PDF export. We compress client side, we force category tagging so the model knows what it is looking at, and we have explicit instructions for the engine to flag unreadable images rather than guess. Guessing on a lab value can hurt a patient. We will never let the engine do that.
Doctors are skeptical, and they should be. Every Filipino doctor I have met has been burned by some overhyped health tech product that promised to change their life and ended up adding three clicks to their day. The only way to win their trust is to put the product in front of them, let them break it, and fix what they break faster than any competitor could. That is what this sprint is building toward. By April 13 we will have a working demo in the hands of our patient zero doctor. By April 27 we close the first five paying doctors. By September we bring it to conference.
Philippine data privacy compliance is not optional. We are building with Data Privacy Act readiness from day one. Encryption at rest, Row Level Security enforcing clinic level isolation, audit logs on every SOAP generation, image upload, and finalization. We are not waiting until we are forced to be compliant. We are building as if we already are.
What the North Star looks like
The SOAP engine is Layer 1 of a four layer vision.
Layer 1 is data ingestion. What you just read.
Layer 2 is clinical decision intelligence. Lab trend analysis across visits. Disease risk prediction tuned to the Filipino population. Polypharmacy flagging across doctors who do not know about each other. A longitudinal health score that tells a doctor in one glance whether their patient is getting better or worse.
Layer 3 is patient facing. A personal health passport accessible by QR code anywhere in the Philippines. A monthly health report card in plain Tagalog. A medication reminder system that actually works because it knows the patient’s full history, not just the prescription from today.
Layer 4 is population health. Barangay level disease prevalence maps. Real time outbreak detection. A DOH partnership layer that gives the Philippine government a view of the country’s health it has never had before.
This will take years. This sprint is one week. But every line of code we write this week is load bearing for the layers above.
What you can do
If you are a Filipino doctor and you want to be one of the first to try MedixAI, reply to this post or reach us at agent@getmedixai.com. We are personally onboarding our first cohort. No sales pitch. We will give you the product, walk you through a real patient workflow, and listen to what is broken.
If you are a developer or a founder building for the Philippines, subscribe. I am publishing the full build log every week. What we ship, what we break, what we learn, what the numbers look like. No gatekeeping. Everything on the table.
If you are a patient who has ever waited two hours to see a doctor for a five minute consultation and walked out with a prescription on a torn piece of paper, this product is for you. Tell the doctors in your life that MedixAI exists.
Next week
Week 2 is the Doctor Dashboard and AI Clinical Brief with severity flagging. The moment a doctor opens their clinic view in the morning, they see every patient for the day, every risk score, every red flag, and every piece of context they need to walk into the exam room ready. I will publish the full feature report next Thursday.
Until then, we build.
No pressure. Just help.
Koleen Baes Paunon Founder, MedixAI — agent@getmedixai.com · https://getmedixai.com
This is Build Log #1 in a weekly series documenting MedixAI in public. Subscribe to follow the build, ship by ship, week by week, until every Filipino clinic is paperless and every Filipino has a health record that travels with them for life.