Inspiration
In rural Nigeria, a Community Health Extension Worker sees 40 patients a day. Her treatment guideline book is in formal English. Her patients speak Pidgin. She does her best, but "her best" means guessing on drug doses and skipping interactions she's unsure about.
75 million Nigerians speak Pidgin. Zero medical AIs understood them. We asked a simple question: what if the computer spoke her language? That question became EARL AI.
What it does
EARL AI is an offline clinical reasoning engine for Nigerian health workers. Type a symptom in Pidgin or English, "my pikin get hot body", and get an instant answer from Nigeria's official treatment guidelines.
- Understands Nigerian Pidgin — not a dictionary, a full NLP layer
- Knowledge graph — 270 conditions, 890 drugs from NSTG 2022
- Drug interaction checks — 164 curated interactions with severity levels
- Smart follow-ups — asks age, weight, symptoms before answering
- Conversational intelligence — adapts tone, tracks topics, detects urgency
- Runs offline — zero internet after install, works on any 4GB laptop
How we built it
We stopped trying to fine-tune big models and built something different: a graph-based reasoning engine.
- Knowledge graph — Parsed all 270 NSTG 2022 conditions into a traversable graph with symptom indexes, drug indexes, and red flag detection
- Pidgin NLP — Built a normalizer with 264 medical terms and 475 phrase mappings using longest-match-first algorithm
- Clinical reasoner — Assesses severity, detects emergencies, calculates drug doses from patient weight
- Conversational engine — Inspired by Fish Audio's multi-turn pattern: topic tracking, adaptive tone, smart follow-ups
- Dual-model architecture — Qwen 1.5B (fast fallback) + MedGemma 4B (best accuracy), both via llama.cpp
Stack: Python (NLP + orchestrator) · Go (DocReader, <5ms lookups) · llama.cpp (inference) · Pure JSON data (no database)
Challenges we ran into
- The 8GB RAM wall — MedGemma 4B alone uses 5GB. We had to make the graph handle 90% of queries so the model only kicks in for edge cases
- Pidgin is messy — "e dey" means different things in different contexts. We built context-aware normalization, not dictionary lookup
- Zero hallucination or nothing — A medical AI that invents a wrong dose is worse than no AI. The graph answers only from verified guidelines
- Cross-platform on a budget — Windows dev machine, Linux eval box. We cross-compiled Go binaries from WSL
Accomplishments that we're proud of
- 1ms per query — faster than most web APIs, with zero hallucination
- 18MB of RAM — the entire reasoning engine fits in less than a single photo
- 270 conditions, 890 drugs — every answer traces back to an official Nigerian guideline
- Pidgin that actually works — understands "my pikin get hot body and dey vomit" as a clinical query
- Runs on anything — Lite mode: 2GB download, works on a 4GB laptop
- 111 passing tests — every component tested, every edge case caught
What we learned
- Bigger isn't always better — an 18MB graph outperforms a 4GB model on 90% of clinical queries
- Language is the real barrier — the gap isn't computing power, it's that medical AI speaks English while patients speak Pidgin
- Offline is non-negotiable — if it doesn't work offline, it doesn't work in rural Nigeria
What's next for EARL AI
- Expand from 270 to 500+ conditions
- Add voice input in Pidgin
- Support Hausa, Yoruba, and Igbo
- Pre-loaded Raspberry Pi kits for clinics with no computers
- Make it the standard clinical tool for every primary healthcare centre in Nigeria
Log in or sign up for Devpost to join the conversation.