HealthBridge

Offline LLM that answers health questions, explains symptoms and treatments, and guides Nigerians and Africans on when to seek urgent care at the hospital.

Comment

Inspiration

Working on a spatiotemporal cholera early warning system for Ondo State for my final year project in Federal University of Technology Akure(FUTA), Nigeria showed me how deadly the gap between health information and people is. Most Nigerians, especially outside major cities and students take medication without seeking out medical practitioners and have no reliable access to a doctor or basic medical guidance when symptoms first appear. That gap costs lives. I wanted to build something that bridges it, with zero internet dependency and zero recurring cost.

What it does

HealthBridge is an offline AI health assistant that runs entirely on a standard laptop with no internet, no GPU, and no cloud. It explains symptoms and treatments in plain language, describes interventions like oral rehydration therapy for cholera and other diseases, and guides users on when symptoms require urgent hospital care without claiming to replace professional diagnosis.

How i built it

  • Selected Qwen2.5-1.5B-Instruct quantized to GGUF Q4_K_M via llama.cpp
  • Validated RAM footprint (peak 1718 MB , it is well under 8GB limit)
  • Achieved 16.76 tokens/sec generation on CPU-only hardware
  • Built download_model.sh for evaluator reproducibility
  • Benchmarked using adtc-profiler v0.1.0 (participant mode)
  • Add a System Prompt to Instruct the model on what to do so it can provide the instructions step by step.

Challenges i ran into

  • Removing accidentally committed large files (1.1GB model) from git history required git filter-repo and force push
  • Balancing model quality vs RAM budget on 8th-gen i5 CPU
  • Ensuring zero network calls during inference while keeping download reproducible for evaluators
  • Modifying the system prompt to provide the right instructions
  • Internet connections issue; the internet connections was breaking down when downloading the model

Accomplishments that i am proud of

  • We haven't accomplish much yet, we are at the MVP phase

What i learned

Quantization level matters more than parameter count for real-world deployment. A well-quantized 1.5B model outperforms a poorly-configured 3B model on constrained hardware. Offline-first AI is not a limitation it is a design choice that makes tools genuinely accessible and free for people to use anytime. The use of Token Usage to prevent words from getting cut off when the model is giving instructions.

What's next for HealthBridge

The Next thing for HealthBridge is to build an app that will be available and accessible for people.

Built With

Share this project:

Updates

Submission history