lumos.health

Inspiration

Coming from a low income background, David's mom was devastated to find out his little brother has reached stage 4 of the Noma disease , astounded and helpless they couldn't do anything. Only if, David had the ability to found out the early symptoms of his brother's disease through understanding the abnormal signs, maybe things could've been different. With a severe fatality rate of 90% when left untreated and Noma is a gangrenous infection and this is what inspired us to build Lumos.health.

What it does

Lumos.health is an all inclusive web app and which can be integrated into your daily messaging environment, which enables CHWs in sub-Saharan Africa to screen for Noma without requiring specialist training. The worker will take a picture of the child's face focusing on the nose and mouth region and send it through a designated platform. Within 8-12 seconds, the platform returns an AI-generated triage decision (Urgent, Refer, Monitor, or Healthy), a clinical note, and a referral to the nearest capable clinic. The system also gives out a 3D render of the child's mouth to depict which areas of the mouth has the infection. Simultaneously, the system acts as a persistent surveillance network; if it detects an outbreak cluster (e.g., 3+ cases within a 10km radius), it automatically fires real-time alerts to health authorities.

How we built it

We built an offline-capable PWA using React 18, Vite, and Tailwind CSS. To accelerate our development timeline, we relied heavily on Cursor, Antigravity, and GitHub Copilot to scaffold the components and API routes while using Express.js/TypeScript backend orchestrated over a Supabase PostgreSQL database equipped with Row-Level Security. The core logic runs on 4 Virtual machine Daedalus swarm utilizing Claude Haiku 4.5. We structured this as a sequential pipeline: VM-1 (Vision Analysis): Examines the photo against WHO staging criteria. VM-2 (Clinical Reasoning): Combines visual data with child's metadata to generate a clinical note. VM-3 (Referral Routing): Queries the database to find the nearest clinic and writes a localized referral note. VM-4 (Surveillance): Runs continuously, calculating Haversine spatial clustering to detect outbreaks. We used Kimmy K2 as the LLM to conduct extensive research on the disease to meticulously map out the clinical reasoning parameters required for the AI pipeline.

Challenges we ran into

Extrapolating real-time data from dense medical research to train and prompt our models required a lot of trial and error to get the clinical reasoning just right. Importing and configuring the 3D mouth model using React Three Fiber was tricky. We had to ensure it rendered smoothly on browsers while accurately representing the specific elements within the mouth that CHWs need to look for. Managing the usage case when a photo is scanned, requires orchestrating 4 distinct VM sub-agents. Passing the payload sequentially without timing out, especially while managing intermittent connectivity fallbacks, was a major architectural hurdle.

Accomplishments that we're proud of

One of our massive achievements, built on an innovative visualization feature where we take a standard 2D image scan of the oral cavity and scale it into an interactive 3D mouth model. Implementing this using React Three Fiber and GLTF models to render smoothly on web browsers while accurately mapping and identifying the specific pathological elements within the mouth was rewarding. Engineering a complex, 4-VM Dadalus compute swarm to run our AI pipeline. Managing the state, handoffs, and concurrent execution across the vision,clinical reasoning, referral, and surveillance agents without dropping requests was a cost effective win.

What we learned

Building Lumos.health taught us how to seamlessly bridge the gap between complex AI infrastructure (managing warm VM pools and multi-agent systems) and low-resource edge environments (offline-first PWAs). We also learned the profound importance of spatial data in public health—realizing that a single data point is a diagnosis, but aggregated spatial data is an outbreak prevention tool.

What's next for Lumos.health

We are looking to scale our product to cater for all diseases that have any external patterns. This can be easily integrated with our pipeline. We would only have to change the factors for each disease.