Gemma 4 Hackathon

Globis Edge Offline caseworker companion [Powered by Gemma 4]

🔐 Secure multimodal intake 📱 Offline on Raspberry Pi ⚖️ PRIMES-aligned 🤝 Human-centered design

📋 Proof of Work Read Kaggle Writeup 1,500-word technical deep-dive 🏆 Main Submission View on Kaggle ⚡ Live Demo 1 min walkthrough 📊 Kaggle Notebook Implementation guide 💻 GitHub Repo Full source code

TL;DR

Offline intake that turns field notes into usable case records.

Capture IDs, notes, and spoken testimony

Structure Clean records directly on-device

Protect Flag conflicts with zero cloud dependency

Build It Yourself

Drag the hardware pieces onto the workbench. The Raspberry Pi, Netac 500GB SSD, USB-C power supply, app screens, and light connection are rendered from the exact artwork. Scroll or pinch to zoom, then drag empty space to pan like a map.

1. Assemble Hardware

Drag components onto the matching workbench targets.

2. Deploy Code

System boots and models load into memory.

3. Connect Phone

Caseworker connects over local WiFi.

4. Process Intake

Inputs are captured, checked, and summarized.

⚡ Offline field intake

What is Globis Edge?

Globis Edge turns IDs, notes, and spoken testimony into structured intake records directly on a Raspberry Pi 5, so teams can keep working when cloud access is unreliable.

📥 Capture messy inputs Audio, OCR, handwritten notes, and typed caseworker context.

🧭 Structure the case Turns fragmented information into reviewable intake records.

🛡️ Keep humans in control Flags conflicts and blocks sensitive fields before final review.

Caseworkers Interpreters Protection staff Low-connectivity field settings

Edge Intelligence Metrics

Drag the dial to compare the verified baseline, what Globis Edge can detect, and what that means in the field.

Field kit run

3 – 5%

UNHCR audit baseline · verified

About 1 error per 30–40 intakes

OIOS 2024/056 found critical registration issues in proGres, with around 700,000 refugee records carrying errors across 2022–2023.

Source: UNHCR OIOS 2024/056 audit report.

70 – 80%

Globis Edge mismatch detection

Catches key discrepancies during intake

Synthetic Scenario A & B testing showed detection across name mismatches, birth dates, family composition, and origin discrepancies.

Field deployment validation is still pending.

11–12s

End-to-end on Raspberry Pi 5

Real hardware, CPU-only runtime

Processes the intake flow locally on Raspberry Pi 5 hardware, from multimodal inputs through structured results.

Measured on the actual edge deployment setup.

$315

Approximate hardware cost

Low-cost field kit deployment

Estimated cost for the Raspberry Pi 5 field kit in USD, making the setup more realistic for constrained environments.

Hardware estimate for the current prototype build.

3 – 4 / 20 – 30

Errors prevented per month

Typical camp impact at two scales

Roughly 3–4 errors avoided monthly in a 150-intake camp, or 20–30 in a 1,000-intake camp.

Impact estimate based on the UNHCR audit baseline.

Slide left / right Baseline

In practice, UNHCR audits show that critical registration errors can take 8+ months to discover and correct during verification. Globis Edge is designed to surface them during intake, before the harm compounds.

8+ monthsaudit discovery lag

CPU-onlyruns on Raspberry Pi 5

$315approx. field kit cost

Explore the field kit

Core Capabilities

A quick, filterable map of what Globis Edge does: collect intake inputs, check conflicts, protect sensitive data, and run fully offline.

🎯

Inference

Tiered Inference

Fast Scout pass for translation and lightweight checks, then Analyst pass for deeper multimodal synthesis.

2-stage local reasoning

🎤

Intake

Multimodal Intake

Audio testimony, identity documents, OCR text, and caseworker notes move through one structured pipeline.

Audio · OCR · Notes

⚠️

Review

Conflict Resolution

Flags mismatches across sources and keeps sensitive decisions human-reviewed instead of auto-resolved.

Human in the loop

🛡️

Safety

Constitutional Audit

Dual-pass verification blocks disallowed fields and fails closed before unsafe records can proceed.

Rule + reasoning checks

🔒

Privacy

Encrypted Persistence

SQLCipher storage, value-masked logs, and append-only quarantine keep sensitive data contained.

Encrypted by default

📱

Edge

Edge Deployment

Runs locally on Raspberry Pi 5 hardware with attached storage and no cloud dependency.

Offline field mode

⚡

Model

Powered by Gemma 4

Multimodal model support for local synthesis, tool use, and structured outputs on constrained devices.

Gemma 4 submission

Explore Globis Edge

Navigate the project like a field map — watch the story, test the demo, inspect the build, or dive into the technical submission.

🧭 Choose your route

Explore by path

Start with a quick preview, follow the full journey, or jump straight into code, documentation, and the Kaggle submission assets.

7 resources Video + code + docs Fast to skim

📹

Full Project Overview

Watch the complete narrative walkthrough from the caseworker’s perspective and understand the end-to-end intake journey.

Story walkthrough

⚡

Quick Demo

Open the 1-minute demo for the fastest look at the interface, workflow, and on-device intake experience.

Fastest entry point

💻

GitHub Repository

Browse the full source code, project structure, tests, and supporting implementation files.

Source code

📚

Full Documentation

Review the PRD, architecture notes, ethics framing, and Raspberry Pi deployment guide.

Docs + deployment

📄

Kaggle Writeup

Read the full technical submission covering the problem, design choices, safety considerations, and impact.

Deep dive writeup

📊

Kaggle Notebook

See the implementation walkthrough, structured notebook flow, and practical integration guide.

Notebook guide

🏗️

Architecture & Workflow

Open the visual system map for the architecture, data flow, and how each component connects together.

Visual system map

Proof of Work / Project Report

1,500+ word technical writeup covering problem, architecture, five hero capabilities, and responsible AI implementation with Gemma 4

📋

1,500+ Word Technical Writeup

Complete Kaggle submission covering problem, architecture, five hero capabilities, challenges, and responsible AI implementation

What's Inside:

The Problem: Caseworkers at Adré, Chad process 40+ intakes daily with fragmented data, no conflict detection
Architecture: Tiered Gemma 4 inference (Scout 2B + Analyst 4B), multimodal intake, cross-modal conflict resolution
Five Hero Capabilities: Tiered inference routing, multimodal ingestion, conflict detection, constitutional auditing, dignity loop
Edge Reality: Runs on Pi 5 (8GB RAM), ~11-12s end-to-end latency, zero cloud dependency
Challenges & Solutions: SQLite vulnerability mitigation, race condition handling, quarantine signaling, protected field logging
Why Gemma 4: Multimodal, native function calling, scale-invariant instruction-following, fits 8GB hardware
Safety Guarantees: Rule-first auditing, fail-closed design, synthetic data only, value-masked logging

View Full Writeup on GitHub →

Project Walkthrough

Exploring Globis Edge architecture, design decisions, and impact

local hello terminal

> _

LinkedIn · quick hello →

Nada Khas

SOFTWARE ENGINEER 👩‍💻 · CANADA 📍

Building systems that reduce friction and improve visibility. I work with Python, TypeScript, SQL, and modern platforms. With Globis Edge 2.0, I applied systems-thinking to a humanitarian challenge: responsible, edge-first AI for low-resource settings where accessibility matters most.

Python TypeScript SQL FastAPI React Systems Design

Get in touch →