ai · 2026
Keiko
A private NCLEX study companion built for one nursing student — the model runs on my desk, the app lives on her phone.
Keiko is a private study platform I built end-to-end for my girlfriend, a nursing student prepping for the NCLEX. It ingests her lecture recordings, PDFs, and slides, then generates NCLEX-style practice questions with paragraph-length rationales pulled straight from her own materials. Everything runs on a single machine on my desk — the language model, the vector store, the transcription pipeline — so nothing about her studying ever leaves the house. The interface is cozy and mobile-first, palette tuned protan-safe because that's what she needs, and it lives behind a Cloudflare tunnel so she can open it anywhere. It's a gift, but it's a real product too — shipping updates every week, hand-tuned to the way she actually studies.
- Local LLM only — Ollama + Qwen 3 on the home GPU, zero cloud dependency
- NCLEX-style MCQ and Select-All-That-Apply with paragraph-length rationales sourced back to her own materials
- Self-hosted RAG pipeline — Marker for documents, WhisperX for lecture audio, pgvector for retrieval
- Cozy, protan-safe design system — palette and correctness cues tuned around one user's actual vision
- Installable PWA behind Cloudflare Access — private URL, email-PIN gate, JWT-verified at the origin
- Next.js 15
- FastAPI
- PostgreSQL + pgvector
- Ollama
- Cloudflare Tunnel
- Docker Compose
The palette avoids red-green pairings — she has protanomaly, so blue means correct and amber means wrong. Every design decision goes through the same test: does it help her tomorrow morning?
