snflwr.ai wraps Open WebUI with a FastAPI backend that enforces 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. It runs entirely on your hardware — no cloud accounts, no data leaving your network.
Why snflwr.ai?
Runs Offline
No cloud, no accounts, no data
leaving your network. Deploy on a
USB drive for complete physical
data control.
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.
Parent Dashboard
Real-time monitoring of every
conversation. Safety incident
alerts, usage analytics, and
full chat history review.
K-5 through 12th Grade
Age-adaptive filtering per child
profile. Content rules tighten for
younger students and relax
appropriately for older ones.
Enterprise Ready
PostgreSQL, Redis, Celery,
Prometheus/Grafana, horizontal
scaling, and COPPA/FERPA audit
trails for school districts.
Encrypted Everything
AES-256 at rest (SQLCipher),
TLS 1.3 in transit, Argon2id
password hashing. PII is never
stored in plaintext.
How It Works
Student Student
│ ▲
▼ │
Open WebUI ──► FastAPI Backend ──► Safety Pipeline ──► Ollama ──► Response
│ Filtered
├─ 1. Input validation
├─ 2. Unicode normalization
├─ 3. Pattern matching
├─ 4. LLM classification (optional)
└─ 5. Age-adaptive rules
All AI inference runs locally via Ollama. The safety pipeline sits between the user and the model — there is no path around it.
Screenshots
Chat Interface
Students interact with a polished AI tutor
Parent Dashboard
Monitor conversations and safety incidents
Setup Wizard
Interactive installer detects your hardware
Create Parent Account
Set up your parent credentials
Set Up Child Profile
Configure age-adaptive filtering per child
Quick Start
Prerequisites: 8 GB RAM recommended (4 GB minimum) · 10 GB free disk · Docker Desktop
# Linux / macOS
chmod +x setup.sh start_snflwr.sh
./setup.sh
# Windows
.\setup.bat
The installer creates a virtual environment, installs dependencies, sets up Docker and Ollama, pulls an AI model sized to your hardware, generates credentials, and writes .env. Then start:
./start_snflwr.sh # Linux / macOS
START_SNFLWR.bat # Windows (double-click)
Open http://localhost:3000 — the startup script launches everything and opens your browser.
Run python install.py directly, then ./start_snflwr.sh.
Double-click Start Snflwr in the USB root. Platform-specific launchers (.bat, .desktop, .command) detect your OS automatically.
Run once: Set-ExecutionPolicy -ExecutionPolicy RemoteSigned -Scope CurrentUser
Or use START_SNFLWR.bat instead.
Deployment
| Scenario | Command | What you get |
|---|
| Family / USB | ./setup.sh && ./start_snflwr.sh | SQLite, AES-256 encryption, fully offline, no Docker needed |
| Home Server | ./deploy.sh | Docker, auto GPU detection, persistent services |
| School / Enterprise | enterprise/build.sh | PostgreSQL, Redis, Celery, Prometheus, Grafana, COPPA/FERPA audit |
./deploy.sh # handles secrets, GPU, images, model, browser
./deploy.sh --stop # stop all services
./deploy.sh --update # pull latest updates
./deploy.sh --logs # tail logs
./deploy.sh --model gemma4:12b # use a smaller model
enterprise/build.sh
docker compose -f docker/compose/docker-compose.yml up -d
See enterprise/README.md for the full guide.
Configuration
The installer generates a .env with all required settings. Key variables:
DB_TYPE=sqlite # or: postgresql
OLLAMA_DEFAULT_MODEL=snflwr.ai-31b # the certified tutor; written by deploy.sh
DB_ENCRYPTION_ENABLED=true
JWT_SECRET_KEY=<auto-generated>
DB_ENCRYPTION_KEY=<auto-generated>
AI Models
The user-facing chat model is always snflwr.ai — what kids see in the
Open WebUI dropdown. It is built locally by the install / deploy scripts as a
base model whose size the installer chooses from your hardware. deploy.sh
auto-selects by system RAM and tops out at gemma4:e4b (it won the June
2026 tutoring bake-off); systems below ~14 GB RAM / ~6 GB VRAM are unsupported
and the install refuses rather than downgrading. Kids
never see the raw base-model tag.
Backbone capability is a function of your hardware, not of "home vs
enterprise." Any box with enough VRAM — a high-end home rig or an enterprise
node — can opt up to the larger gemma4:31b backbone; a modest box of either
kind stays on e4b. The one floor that matters is VRAM co-resident with the
~5 GB llama-guard3:8b safety classifier (always loaded alongside the tutor —
that's why each row's VRAM floor exceeds the model's own size):
| What | Model | Weights | Requirement | Notes |
|---|
| Tutor | snflwr.ai-31b (built on gemma4:31b) | ~19 GB | ~22 GB VRAM | The only backbone with a sealed tutoring run. A box that cannot hold it does not tutor — the serving plan refuses rather than serve an unmeasured model. |
| Safety classifier | llama-guard3-cpu | ~5 GB | RAM, not VRAM | CPU-pinned on purpose. On the card it was evicted mid-request and failed closed, blocking children on harmless questions. Same blob digest as the GPU build; 1.9 s vs 1.8 s. |
| Homework gate / reveal confirm | gemma4:e4b (base weights) | ~10 GB | RAM | Pinned away from the tutor model deliberately: run on the tutor, the confirm inherited its persona, emitted two characters and scored 0% recall on 20 known reveals. |
There is one tutor, not a ladder (2026-09-20). The smaller backbones were
retired from tutoring on measured quality: gemma4:31b scored 88.8 against
e4b's 80.3 overall, and on the same 121 probes cut wrong content 17 → 4,
canned fallback 27 → 5 and refusals 31 → 10. Throughput is the real cost — 31b
serves roughly a quarter of the students per card — and that is not a reason to
teach children wrong content.
Ask what a given box can serve:
python3 scripts/certified_tutor.py # model, base and context window, or exit 3
Retired claim. This table used to gate 31b at ≥26 GB "because it cannot
co-reside with the 5 GB guard". That premise is gone: the guard is CPU-pinned,
and a 23 GB card serves the 31b tutor in production today.
The wrapper bundles the K-12 tutor system prompt (all school subjects — STEM
plus reading, writing, history, civics, and the arts), sampling parameters
(including repeat_penalty to prevent reasoning loops), and safety stop
sequences from models/Snflwr_AI_Kids.modelfile.
To switch the underlying base, re-run ./deploy.sh --model gemma4:e4b
(or set BASE_MODEL in .env.home and re-run ./deploy.sh). The
snflwr.ai wrapper will be rebuilt automatically on top of the new base.
Streaming (optional): set CHAT_STREAMING_ENABLED=true for hold-back
streaming — the tutor's reply is streamed once the safety pipeline has vetted it,
so the first vetted text appears in ~1–2 s instead of after the full response,
without weakening output checking.
Security & Compliance
Compliance status: the items below are COPPA/FERPA-supporting
engineering, not legal certification. The legal documents in legal/
are DRAFT and no business entity is registered yet, so snflwr.ai is not yet
legally launchable as a paid service. See docs/CURRENT_STATUS.md.
| Layer | Implementation |
|---|
| Encryption at rest | AES-256 via SQLCipher |
| Encryption in transit | TLS 1.3 |
| Password hashing | Argon2id (PBKDF2 fallback) |
| Authentication | JWT with secure session management |
| Privacy | All data local — never sent to external APIs |
| COPPA | Parental consent flow, data minimization, automated retention cleanup |
| FERPA | Student record protections, parent/guardian access controls |
| GDPR | Data deletion and export endpoints |
See SECURITY.md for the vulnerability disclosure policy.
Testing
pytest tests/ -v -m "not integration"
3,500+ tests across 100+ test files at 88% coverage — authentication, profiles, safety pipeline, encryption, database, API routes, middleware, WebSockets, caching, and model management.
Monitoring
Enterprise deployments include Grafana dashboards, Prometheus alerting, and Sentry error tracking with COPPA-compliant PII filtering. See MONITORING_GUIDE.md.
Documentation
Contributing
We welcome contributions in safety filter accuracy, multi-language support, edge case testing, documentation, and UI/UX. Please read CONTRIBUTING.md and CODE_OF_CONDUCT.md before opening a PR.
Support
License
Proprietary — see LICENSE. All rights reserved; no use, copying, modification or distribution except under a written agreement with snflwr.ai.
Versions published between 2026-02-27 and 2026-09-18 were released under AGPL-3.0, and copies obtained in that window remain governed by AGPL-3.0.
The Open WebUI frontend has its own license: frontend/open-webui/LICENSE.
Commercial licensing: [email protected]
Acknowledgments
- Open WebUI — Open-source AI interface
- Ollama — Local LLM inference
- Google DeepMind — Gemma model family (gemma4:e4b, the default tutor backbone)
- Meta — Llama Guard safety classifier
- K-12 educators who provided feedback and testing
