Ship Your First On-Device Model This Weekend
3 emails. 3 days. From zero to a fine-tuned model running locally on your hardware. No ML expertise, no Python scripts, no YAML configs. Total time: ~4.5 hours.
Total hands-on time across all 3 days
Per-inference cost after fine-tuning
Lines of ML code required
Emails — one per day, straight to the point
What You'll Build
Day 1
Collect Your Data (2 Hours)
- Define your task: classification, extraction, generation, or vision
- Minimum viable dataset sizes by task type — the real numbers, not academic ideals
- Data collection shortcuts: scrape app logs, synthetic augmentation, manual labeling
- Format your data into instruction/response pairs or conversation format
- Quality checklist: you're ready for Day 2 when every item passes
Day 2
Fine-Tune Your Model (30 Minutes)
- Choose your base model: 1B for mobile, 3-7B for tablet/edge, 7-14B for desktop
- Upload dataset to Ertas and configure training visually — no code, no YAML
- What the training metrics actually mean (loss curves, eval results) in plain language
- Download your fine-tuned GGUF model file
- Quick validation: test 10 prompts and compare to base model
Day 3
Deploy and Integrate (2 Hours)
- Runtime options by platform: llama.cpp (desktop), MLC-LLM (mobile), Ollama (server)
- Quantization guide: Q4_K_M vs Q5_K_M vs Q8 — when to use which
- Integration patterns: REST wrapper, direct library call, background service
- Testing: what 'good enough to ship' looks like for v1
- Ship it — you now have a feature no one can replicate by calling an API
This Course Is For You If...
You're building an app that needs AI features but API costs worry you
You've heard of fine-tuning but it sounds like an ML engineering project
You want your app to work offline or with zero-latency inference
You're a vibecoder who ships fast and wants to add AI that's actually yours
You care about user privacy and want data to stay on-device
You've hit the prompt engineering ceiling and need the next level
Your Weekend Project Starts Now
Day 1 arrives within minutes. By Sunday, you'll have a fine-tuned model running on your hardware.