We’re pleased to invite you to the upcoming EMR4ALL Hackathon, taking place from July 21 to 25. This event is designed to engage contributors across development, implementation, and advocacy with a strong focus on building offline-ready digital health tools for underserved communities.
Developer Track
This is an ideal opportunity for new OpenMRS developers to:
Explore the OpenMRS codebase
Work on beginner-friendly tickets
Learn best practices
Collaborate through hands-on mentorship
If you’re an experienced contributor, we kindly ask you to consider joining as a peer reviewer or mentor to guide newcomers during the hackathon.
EMR4ALL Product Focus
We will be working on:
AI features that operate offline
Sync improvements
Raspberry Pi-based deployments
Field-ready EMR kits
Implementation & Advocacy Tracks
Additional tracks will focus on:
Supporting implementers deploying OpenMRS in rural, low-connectivity areas
Developing outreach and fundraising strategies
Engaging with Ministries of Health and partner organizations
We’ve been experimenting with offline AI features for OpenMRS and made some positive progress. A few months ago, we successfully ran Ollama on a Raspberry Pi in full offline mode. You can check out a demo here: https://www.youtube.com/watch?v=tMbCsbM3p7E
We’ve also integrated DeepSeek to enable querying the OpenMRS database using LLM in natural English. To handle the compute load on Raspberry Pi 5, we added the Hailo AI accelerator, AI Kit - Raspberry Pi Documentation which significantly boosts performance. The results are impressive, even offline, queries run with minimal latency. Where internet is available, responses are faster than most cloud-based AIs we’ve tested!
We’re now working on a friendly UI to allow clinicians and users to easily run queries and generate cohort reports for broader use across deployments. @tendomart can provide more tech details if needed .
While finalising our AI integration, we conducted several tests using different LLMs to achieve faster response times, reduced latency and fewer hallucinations. In offline mode, queries are processed locally on a Raspberry Pi enhanced with a Hailo AI accelerator , offers an accessible way to deliver local, high-performance and power-efficient inferencing.
The result is low-latency processing with high accuracy, particularly when the AI is trained to preform specific domain tasks on regular basis and data is stored locally. This is different when the same computer is querying Cloud-based AI models wich require a stable internet connection to process data to the remote servers.
For example, considering a rural health facility with a 2G internet connection querying an online cloud AI, latency can increase to 2–3 seconds (or more) before the cloud AI server replies. However, when this is done locally, performance can drop to 1–2 seconds with the same LLM structure running offline.
From a cost perspective, an offline AI acceleration setup comprising a Hailo and a Raspberry Pi 5 with 16 GB of RAM costs under $250 USD , a one-time investment while to reach high performance on Cloud AI you need subscription . This represents a significant advantage compared to recurring cloud service fees and the challenges of ensuring consistent internet access in health facilities especialy in remote setting. Furthermore, offline AI offers better control over sensitive health data.
Our goal at @EMR4All is to optimise AI for use in remote or low-resource settings, with our current focus being on primary healthcares in rural communities. @tendomart has been playing an active role in the AI development and we have observed promising results. We hope to present a live demo to the OpenMRS community soon to showcase this advancement.
I totally understand the target of this project for low resource settings. I just wanted to share your experience in regards to the differences in accuracy for locally run AI models compared with the more powerful cloud hosted ones, for this particular use case.
Thank @dkayiwa We had started to deploy a hybrid AI model using a large language model (LLM) hosted on AWS. The intention was to provide advanced AI capabilities, such as clinical documentation generation, decision support and medical summarisation, during periods of limited connectivity, particularly when integrated with OpenMRS. This approach combined Ollana, an offline AI application running on a Raspberry Pi, with a cloud-based LLM to enable more powerful inference. Although AWS was initially used, the architecture was designed to avoid vendor lock-in and support deployment in environments where privacy, data control and flexibility are paramount, such as ministries of health or NGO partners. The LLM was intended to be accessible via secure APIs or as a containerised deployment, synchronising with the local installation as required. However, we eventually paused the cloud-based work on AWS to refocus on enhancing the offline-first solution.
yes performancewise, cloud hosted ,models currently perform better than those hosted on edge devices, contextually speaking Raspi. we have however added AI accelerator chips to the device (FYI this is like the entry device of a 8GB RAM ), response has been faster also considering that the device has been running other containers like Bhamni and OpenMRS 3.0 Distro. This is a tradeoff especially given the costs associated with enterprise hosted models and those power hungry GPU’s.
Currently we are working to improve
Performance
“Callibrating “ i.e experimenting to find the best perfoming models in terms of accuracy(Speed is not so much an issue for average param models (1-7B)
Coming up with better architectures, agentic patterns e.t.c to drive the use cases at hand and others as we go by.
View the agenda and sign up here