Inspiration
Across Nigeria and much of sub-Saharan Africa, the community pharmacist — not the doctor — is where most people actually get healthcare. Pharmacies make up close to half of Nigeria's primary healthcare facilities and handle around 70% of registered care visits. Patients often move between pharmacies with no shared records, no physician oversight, and a pharmacist who has to catch dangerous drug combinations from memory while a queue builds at the counter. We wanted to build something that gives that pharmacist backup — not a doctor replacement, but a reasoning partner.
What it does
PharmaGuard is an AI co-pilot for pharmacists. You type or speak a patient's current medications in plain English or Pidgin, and it checks them against a pharmaceutical knowledge graph in seconds. If it finds a dangerous combination, it doesn't just flag it — it explains the mechanism (shared metabolic pathway, organ load, etc.) in language the pharmacist can repeat to the patient, and suggests safer alternatives.
How we built it
The core is a knowledge graph connecting drugs, metabolic enzymes, and pathways, traversed to find interaction paths between any two drugs. An NLP layer extracts drug entities from free-text or spoken input in English and Pidgin, and an LLM layer turns the raw graph path into a plain-language explanation. Backend runs on Python/FastAPI and Node.js/TypeScript with PostgreSQL, built to stay lightweight enough for a mid-range smartphone on low bandwidth.
Challenges we ran into
Getting reliable parsing of Pidgin and informal drug names was harder than standard medical English. Balancing explainability with speed was another — it's easy to build a black-box classifier, much harder to make it show its reasoning path clearly enough for a pharmacist to trust and repeat to a patient. Working solo also meant every layer — NLP, graph, backend, explanation generation — had to be scoped tightly to fit the hackathon timeline.
Accomplishments that we're proud of
Getting the reasoning engine to trace an actual interaction pathway rather than doing a flat lookup — the system can say why two drugs conflict, not just that they do. That explainability is the whole point: in healthcare, a flag nobody trusts doesn't get acted on.
What we learned
How much of healthcare access in this region already runs through pharmacists, and how little tooling exists to support that role specifically — most health tech targets doctors or hospitals, not the counter where most people actually show up. Also learned a lot about the tradeoffs between knowledge-graph precision and coverage when you don't have a fully curated local drug database yet.
What's next for PharmaGuard
Voice input, wider drug coverage with dosage and kidney-function-aware checks, and a pilot with 3–5 community pharmacies to refine the language model on real Pidgin usage and expand the graph based on what pharmacists actually ask it.