Safe AI for every
classroom and home
Your child talks to AI. You control what it says back.
snflwr.ai wraps Open WebUI with a FastAPI backend enforcing multi-layer content filtering, parental oversight, and encrypted data storage. Every message passes through a 5-stage safety pipeline that cannot be bypassed from the frontend. Runs entirely on your hardware — no cloud accounts, no data leaving your network.
Why snflwr.ai?
Everything you need for safe AI in K-12
Designed for educators, parents, and school districts who need real safety guarantees — not just content policies that can be edited in a browser.
Runs Offline
No cloud, no accounts, no data leaving your network. Deploy on a USB drive for complete physical data control. All AI inference runs locally via Ollama.
Fail-Closed Safety
5-stage content pipeline: input validation, normalization, pattern matching, LLM classification, and age-adaptive rules. If any stage errors, content is blocked — never passed through.
Parent Dashboard
Real-time monitoring of every conversation. Safety incident alerts, usage analytics, and full chat history review. Know exactly what your child is asking.
K-5 through 12th Grade
Age-adaptive filtering per child profile. Content rules tighten for younger students and relax appropriately for older ones — calibrated, not one-size-fits-all.
Enterprise Ready
PostgreSQL, Redis, Celery, Prometheus/Grafana, horizontal scaling, and COPPA/FERPA audit trails. Built for school districts and enterprises, not just home deployments.
Encrypted Everything
AES-256 at rest via SQLCipher, TLS 1.3 in transit, Argon2id password hashing. PII is never stored in plaintext. COPPA, FERPA, and GDPR endpoints included.
Under the hood
How It Works
The safety pipeline sits between the user and the model. There is no path around it.
Student opens chat interface
Students interact with a polished AI tutor interface powered by Open WebUI. No accounts required for students — parents manage profiles.
Message passes through 5-stage safety pipeline
Every message is processed by: input validation → Unicode normalization → pattern matching → LLM classification → age-adaptive rules. If any stage errors, the content is blocked.
- Input validation
- Unicode normalization
- Pattern matching
- LLM classification
- Age-adaptive rules
AI processes locally via Ollama
The message is sent to a local Ollama instance running Qwen or any compatible model. Nothing leaves your machine. GPU acceleration is auto-detected at setup.
Filtered, safe response returned
The AI response also passes through the safety pipeline before the student sees it. Parents receive real-time alerts for any flagged content, with full conversation history in the dashboard.
Open Source · AGPL-3.0
Ready to deploy safe AI?
8 GB RAM recommended · 10 GB free disk · Docker Desktop