Free Cookbook
The On-Device AI Cookbook
10 practical recipes for building on-device AI features. Each recipe tells you exactly what model to use, how much data you need, where to deploy, and how long it takes to ship.
On-device is the right architecture when any of these conditions apply:
Real-time response needed
Unreliable connectivity
Privacy sensitivity
Per-inference cost matters
All Recipes
Mobile Vision
Real-Time Produce Identification at POS
Why on-deviceSub-100ms latency, works without internet at checkout
ModelVision classifier, 1-3B, Q4_K_M
Data needed500-2,000 labeled images of store inventory
Deploy toAndroid tablet at checkout counter
Time to shipWeekend project with existing inventory photos
Privacy Consumer
Privacy-First Document Scanner
Why on-deviceSensitive docs never leave the phone — GDPR by architecture
ModelText extraction, 1.5-3B, Q4_K_M
Data needed300-500 document/extraction pairs
Deploy toiOS/Android mobile app
Time to ship1-2 weekends with sample documents
Audio & Sensor
Industrial Machine Sound Anomaly Detection
Why on-deviceFactory floor has no WiFi, millisecond response for safety
ModelAudio classifier, 0.5-1.5B, Q4_K_M
Data needed1,000+ labeled audio clips (normal vs anomalous)
Deploy toRaspberry Pi or edge device on factory floor
Time to ship1-2 weeks with audio collection
Text & Language
Offline Translation for Field Workers
Why on-deviceWorks in remote construction sites with zero connectivity
ModelTranslation, 1.5-3B, Q5_K_M
Data needed2,000+ sentence pairs in domain vocabulary
Deploy toRugged Android tablet or phone
Time to ship1-2 weeks with bilingual corpus
Field & Industrial
Weld Quality Inspection
Why on-deviceInstant pass/fail at the weld site, no upload delay
ModelVision classifier, 1-3B, Q4_K_M
Data needed500-1,000 labeled weld images (pass/fail + defect type)
Deploy toTablet or phone with camera
Time to ship1-2 weekends with labeled photos
Personalization
Personalized On-Device Autocomplete
Why on-deviceLearns your slang without sending it to a server
ModelText generation, 0.5-1.5B, Q4_K_M
Data needed1,000+ message/completion pairs from user data
Deploy toMobile keyboard integration
Time to shipWeekend project with synthetic data
Mobile Vision
Plant Disease Identification
Why on-deviceFarmers need answers in the field, not in WiFi range
ModelVision classifier, 1-3B, Q4_K_M
Data needed1,000+ labeled plant images per disease type
Deploy toMobile phone app
Time to ship1-2 weeks with image dataset
Privacy Consumer
On-Device Content Moderation for Kids' Apps
Why on-deviceText never leaves the device — required for COPPA compliance
ModelText classifier, 0.5-1.5B, Q4_K_M
Data needed500-1,000 labeled examples per category
Deploy toiOS/Android app runtime
Time to shipWeekend project with labeled data
Agriculture
Conveyor Belt Fruit Grading
Why on-deviceReal-time sorting at processing speed, no network round-trip
ModelVision classifier, 1-3B, Q4_K_M
Data needed1,000+ labeled fruit images (size, ripeness, defects)
Deploy toEmbedded camera on sorting line
Time to ship1-2 weeks with production line images
Audio & Sensor
Voice Command Recognition for Warehouse
Why on-deviceGloved workers can't touch screens, needs offline + fast
ModelAudio classifier, 0.5-1.5B, Q4_K_M
Data needed500+ recordings per command in warehouse noise
Deploy toHandheld scanner or headset with edge compute
Time to ship1-2 weeks with command recordings
Start Fine-Tuning Today
Pick a recipe, collect your data, and fine-tune on Ertas. No ML expertise needed — upload, train, download GGUF, deploy.