Phase

Inspiration

The sole female member of our team inspired us to find an all-in-one solution to help PMOS / PCOS (Polyendocrine Metabolic Ovarian Syndrome / Polycystic Ovary Syndrome) patients track their symptoms, prepare for appointments, and find answers. Up to 70% of women with this condition go undiagnosed, and patients often spend 2+ years bouncing between doctors who dismiss their symptoms or blame them on unrelated issues. We wanted to build a platform that provides AI-powered, empathetic insights backed by peer-reviewed clinical research and real physician feedback, giving patients the clarity and data they need to advocate for themselves.

Our difference

Women’s health is already under-researched, and PMOS even more so. Most healthcare services incorrectly identify PMOS as other diseases due to the variety of symptoms that can lead to diagnosis, and while symptom trackers do exist for PMOS, none are able to provide the personalized treatment options that PHASE does. Many apps simply save the information put in by users without explaining how the data is used and don't give feedback beyond generic, textbook-based responses. Meanwhile, PHASE takes into account the symptoms inputted by the user to build off of available data and provide the best solutions for their specific experiences.

What it does

PHASE has three main features: Track, Insights, and Ask.

1. Track

Allows users to log and monitor their health across several key areas:

Daily symptoms: Tracking severity and notes for acne, fatigue, hirsutism (excess hair), pelvic pain, mood swings, and sleep.
Cycles & Medications: Logging cycle length, flow, and daily medications or supplements (like Metformin, Inositol, or birth control), including dosages and side effects.
Lab markers: Storing bloodwork numbers (such as Fasting Insulin, HbA1c, Free Testosterone, and DHEA-S) alongside normal ranges.

All logs are securely synced to a managed Delta Lake table (user_health_records) using the Databricks SQL Statement API, keeping patient data organized, scalable, and private.

2. Insights

Takes everything the user logs and turns it into clear, actionable visuals:

Trend graphs: Lets users spot correlations between their medications, cycle days, and symptom flare-ups over time.
Phenotype Quiz: Users can take a clinical assessment that categorizes their symptoms into one of four recognized phenotypes: Insulin-Resistant, Inflammatory, Adrenal, or Post-Pill. This helps predict root drivers, suggests personalized nutrition and lifestyle habits, and prevents misdiagnosis.
Doctor Visit Summary: Generates a clean, exportable one-page health report that users can print or show on their phone during a 15-minute appointment so their doctor has all the context at a glance.

3. Ask

Waiting months for a specialist visit often leaves patients anxious and searching online through conflicting forum posts. Our Ask feature gives users personalized, evidence-grounded answers.

This isn't just a generic AI prompt wrapper:

Clinical Literature RAG: We query a Databricks Vector Search index loaded with peer-reviewed medical guidelines (including ESHRE/ASRM consensus papers and endocrinology journals) to attach exact study names, publishers, and excerpts directly to the answer.
Doctor-in-the-Loop Feedback: We built an expert review system where clinicians (OB-GYNs, endocrinologists) can review answers, rate them on accuracy, empathy, and safety, and write approved corrections. These audits are saved in a Delta Lake table (expert_response_reviews).
Self-Improving Prompting: When a user asks a question, our backend checks for past doctor-approved corrections with high accuracy scores (4+ out of 5) that match the question's keywords. The verified doctor's correction is injected directly into the prompt for Databricks Model Serving (running Llama 3.3 70B Instruct). This means when an expert fixes a mistake, the AI learns from it and doesn't repeat it.

How we built it

We built PHASE with performance, reliability, and privacy at the center:

Frontend & Backend: Built with Next.js 16 (App Router), React 19, and TypeScript for fast, server-side data handling.
Authentication & Privacy: Powered by Clerk so all user logs are scoped strictly to their private account. No patient data is ever made public or used to train external public models.
Data & AI Layer (Databricks):
- Databricks Serverless SQL Warehouse: Executes SQL statements (CREATE TABLE IF NOT EXISTS, MERGE INTO, SELECT) on Delta Lake tables.
- Databricks Vector Search: Runs fast hybrid searches across our indexed medical literature.
- Databricks Model Serving: Serves our Llama 3.3 70B clinical model.

Challenges we ran into

Connecting Databricks with Next.js: None of us had used Databricks before this hackathon! Learning how to authenticate and query Databricks with REST endpoints (the SQL Statement API, Vector Search, and Model Serving) from server-side Next.js routes took hours of trial, error, and debugging.
Preventing AI Hallucinations: Medical topics require extreme care. We spent a lot of time fine-tuning our system prompts and structuring the context so the AI only references the user's actual journal dates and verified research papers.
Team Coordination: With different workspaces and setups, we had to carefully design mock fallbacks and modular code so teammates could test UI features locally even when the frontend was still being developed.

Accomplishments that we're proud of

Pushing through hours of roadblocks until our live Databricks SQL, Vector Search, and Model Serving endpoints all communicated seamlessly with our web app.
Building a real Human-in-the-Loop review loop where verified clinician feedback actively guides future AI answers.
Delivering a full-stack, cohesive product that feels calm and supportive rather than overwhelming or clinical.

What we learned

How to design a modern Lakehouse architecture with Databricks: managing Delta Lake tables, querying through the SQL Execution API, and deploying Vector Search indexes.
How to use in-context learning with verified examples to make LLM responses safer and more consistent.
How to coordinate a full-stack project using Git under a tight hackathon deadline, and how to pitch our solution clearly with great feedback from mentors.

What's next for PHASE

Community Boards: Adding an anonymous, phenotype-matched community space where users can discuss symptoms and share experiences safely.
MLflow Fine-Tuning: Using MLflow to track expert ratings over time and fine-tune smaller, specialized models directly on physician-verified response pairs.
Live PubMed Syncing: Setting up automated pipelines to ingest new clinical trials and research papers into the Vector Search index as soon as they are published.