TibaEdge

Inspiration

Across many African communities, frontline health workers operate with limited internet access, overstretched facilities, and incomplete access to up-to-date clinical reference materials. Most AI health tools depend on cloud APIs, making them expensive, unreliable during connectivity outages, and unsuitable for sensitive patient information.

TibaEdge was inspired by a simple question: can useful, safety-conscious medical guidance run entirely offline on the affordable laptops already available in clinics and community health centres?

What it does

TibaEdge is a safety-first, offline clinical decision-support assistant for frontline African health workers.

The first prototype is designed to help a health worker:

• Record a patient's symptoms, history, and basic vital signs.
• Identify warning signs that may require urgent referral.
• Retrieve relevant information from trusted local medical references.
• Suggest follow-up questions for the health worker to consider.
• Explain recommended next steps in clear, patient-friendly language.
• Continue working when internet access is unavailable.

TibaEdge is not intended to replace a doctor, make a definitive diagnosis, or independently prescribe medication. It supports trained health workers while keeping clinical decisions under human control.

How we are building it

TibaEdge is being designed around a compact language model converted to GGUF format and executed locally through llama.cpp.

The planned architecture combines:

1. A quantized language model optimized for an 8 GB laptop.
2. Local retrieval over curated healthcare guidelines.
3. Rule-based checks for critical warning signs.
4. Clear citations showing which local reference informed an answer.
5. A lightweight interface that works without cloud services.
6. Local-only processing so sensitive information does not leave the device.

The model will be profiled for response quality, tokens per second, peak memory usage, first-token latency, and CPU temperature using the official ADTC profiler.

Challenges

The central challenge is balancing medical usefulness with strict hardware and safety constraints. A larger model may provide stronger answers but consume too much memory or respond too slowly. A smaller model is more practical, but requires careful prompting, retrieval, domain adaptation, and safety guardrails.

Other challenges include reducing hallucinations, presenting uncertainty honestly, selecting trustworthy medical references, and ensuring that urgent warning signs are never hidden inside a long generated response.

What we have learned

Building for constrained hardware changes the definition of a good AI product. Accuracy matters, but so do memory efficiency, response time, privacy, explainability, and the ability to work reliably without connectivity.

For a medical system, knowing when not to answer and when to recommend urgent human intervention is just as important as generating a fluent response.

Current stage and what's next

TibaEdge is currently at the early-prototype stage. The user workflow, safety boundaries, offline architecture, and benchmarking plan have been defined.

The next steps are to select and benchmark suitable base models, build the local retrieval pipeline, validate the warning-sign rules with qualified healthcare professionals, and test the complete system on the ADTC standard laptop profile.

Built With

• face
• generation
• gguf
• github
• gradio
• hugging
• language
• large
• learning
• llama.cpp
• machine
• models
• natural
• processing
• python
• quantization
• retrieval-augmented
• sqlite
• ubuntu