Medeté

Your Pocket AI Tutor for Global Medical Entrance Exams

Comment

Inspiration

In many parts of the world, aspiring doctors fight two invisible battles: poor connectivity and lack of personalized support. Coaching centers cost a fortune, and cloud AI tutors demand reliable internet — a privilege, not a guarantee. We asked:

“What if every smartphone could become a truly private, always-available AI medical coach?” That question sparked Medeté — the world's first fully on-device AI tutor for global medical entrance exams.

What it does

Medeté is an offline, privacy-first AI medical tutor built for and by medical aspirants.

It helps students preparing for NEET, MCAT, and GAMSAT generate:

Personalized flashcards

Practice quizzes

Socratic-style tutoring anytime — even without internet

Core features:

AI-powered flashcard & quiz creation

Chat-based tutor trained on medical reasoning chains

Local semantic search over compressed knowledge base

Smart progress analytics: daily streaks, topic mastery heatmaps

Ghost-prompting for vague queries

Shadow-testing to reduce hallucinations

On-device LoRA fine-tuning for local learning

⚡ All runs on-device → ~3s response time even on mid-range phones.

How we built it

Layer Stack / Tools Frontend Flutter (Dart) – cross-platform mobile AI Quantized Llama-3-8B-Instruct (4-bit GGUF) via llama.cpp Search Local Faiss vector store Storage SQLCipher for encrypted analytics Backend Supabase Functions (non-critical tasks) Bridge Dart/C++ via FFI CI/CD GitHub Actions, automated builds & deploys (took the help of bolt ai too) Other unique pieces:

Ghost-prompting engine (Python)

Shadow-testing harness

Federated fine-tuning buffer

Challenges we ran into

Keeping large models responsive on mid-tier phones

Medical answer accuracy without server validation

App Store restrictions on AI model loading

Designing privacy-first federated learning

Local storage & token expiry offline

Balancing battery, performance, and heat

Accomplishments that we're proud of

The world is flat. Achieved sub-3s local inference! ⚡

Hybrid memory: session + episodic + semantic

Shadow-testing layer to auto-detect hallucinations

On-device LoRA personalization without data leaks

Stunning Flutter UI & dark mode 🌙

<500MB install size

HIPAA compliance + differential privacy in analytics

What we learned

Quantization & hardware acceleration change everything

Federated learning works — if you buffer carefully

Students love voice input 🎤

UX details (dark mode, streaks) boost engagement

Shadow-testing can meaningfully catch hallucinations offline

What's next for Medeté

Smart revision planner

Multi-modal learning (diagrams, videos, audio)

Federated social study groups

Marketplace plugins (educator-uploaded question banks)

Behavioral analytics to spot burnout

Expand to USMLE, PLAB, AIIMS

Counselor dashboards

Big vision: build the world’s first ==truly offline, private, AI-powered learning companion== — starting with medical exams, then helping every learner, everywhere.

Built With

Share this project:

Updates