GlycoShelter

The problem

Many unhoused individuals struggle with diabetes and metabolic health. Congregate shelters and public food banks often serve carbohydrate-heavy, highly refined starches (white bread, pasta, processed sugars) because they are cheap. Unhoused diabetics must navigate this landscape blind, unable to predict if their dinner will cause a glycemic emergency.

What it does

GlycoShelter is a mobile-first geographic directory of San Francisco shelter facilities.

• Menu Ingestion: It ingests raw, unstructured daily menus (uploaded as text or PDFs) from local shelter coordinators.
• GenAI Analysis: Running on a DigitalOcean Serverless Function, it processes the text through a hosted Llama-3-70B model to perform zero-shot glycemic analysis.
• Metabolic Scoring: It outputs structured glycemic classifications (LOW, MEDIUM, HIGH) and isolates high-risk items.
• Visual Mapping: The map populates clean, color-coded, glowing pins allowing users to visually spot metabolically safe meal locations near them.

How we built it

• Frontend: A responsive Tailwind CSS dashboard utilizing Leaflet.js map instances. We injected a custom CSS filter stack to create a high-contrast dark theme, focusing on visual legibility for individuals experiencing diabetic vision fatigue.
• Datasets Used: We pulled location and capacity data from the San Francisco Department of Homelessness and Supportive Housing ("Shelter Occupancy Rate and Total Temporary Shelter Beds") and merged it with food services data from the SF "Registered Business Locations" directory.
• Backend & AI Pipeline: We designed a DigitalOcean Serverless Function template that utilizes pdf-parse to convert unstructured menu PDFs into plaintext. This text is passed to DigitalOcean's GenAI Platform using Llama-3-70B to return a standardized JSON metabolic safety payload.

Challenges we overcame

Handling the unpredictable format of daily menus (handwritten notes, basic tables, inconsistent text) was a challenge. By utilizing a high-parameter LLM (Llama-3-70B) via DigitalOcean, we were able to run reliable zero-shot extraction that normal regex parsers would completely break on, guaranteeing a clean JSON payload for our map frontend.

Built With

• and
• css
• department
• digitalocean
• francisco
• homelessness
• llama-3-70b
• of
• san
• serverless
• supportive
• tailwind