~/projects/behaveMar 2026 — Present
BeHave
On-device behavior design app running Gemma 3n (INT4) fully offline — 10–30 tok/s on mid-range Android devices, zero cloud dependency.
- Kotlin
- Jetpack Compose
- MVVM
- LiteRT-LM
- Work Manager
- Room
An expand on MVVM layering, Room persistence, how behaviors are prompted, WorkManager scheduling.
Challenges & solutions
Quantized LLM outputs are noisy and break strict JSON parsers. Built a JSON parser with retry logic and a rule-based fallback path, reaching >95% parse success rate.
Outcomes & impact
Full local LLM inference on mid-range devices with no API keys, no network calls, no per-token cost — privacy-first by design.
Working on something similar?