OoJae/naija-twin

# Naija-Twin **Your AI twin that thinks Naija.** Most LLM recommendation systems treat user modeling and recommendation as two separate pipelines. Naija-Twin collapses them into one. A single persona memory brain drives both a User Simulator (Task A: reviews and ratings) and a Recommender (Task B: cold-start, cross-domain, multi-turn). Every improvement to the persona lifts both tasks at once. The system speaks authentic Nigerian English across five registers and is evaluated on a 198-item Nigerian-fit benchmark built from scratch. **Live demo**: naija-twin-web.vercel.app **Live API**: naija-twin.onrender.com **Paper**: paper/paper.html — also served at naija-twin-web.vercel.app ## Architecture: Twin-Loop ``` +---------------------+ | Persona Memory | | (Semantic + | | Episodic + | | Procedural) | +--------+------------+ | +--------------+--------------+ v v +-----------------+ +-----------------+ | Task A: | | Task B: | | User Simulator | | Recommender | | (Reviews + | | (Cold-start, | | Ratings) | | Cross-domain, | | | | Multi-turn) | +-----------------+ +-----------------+ | | +--------------+---------------+ v +---------------------+ | Reflection Loop | | (Async critique, | | faithfulness) | +---------------------+ ``` ## Headline Metrics | Metric | Value | Target | Baseline | |--------|-------|--------|----------| | NDCG@10 (persona-anchored split) | **0.471** | >= 0.45 | P5 zero-shot ~0.32 | | RMSE (rating prediction) | **1.080** | = 90% | Untuned MiMo ~73% | | Nigerian-factor mention rate | **70%** | >= 40% | Untuned MiMo ~45% | | Register accuracy (80 reviews) | 64% | >= 70% | - | | ROUGE-1 F1 (review generation) | 0.205 | 0.18-0.25 | - | | MAE | 0.778 | - | - | | ECE (calibration) | 0.009 | - | - …