Lumina-Agent

Champion · Rank 1
Lumina-Agent is a data-centric, memory-aware end-to-end voice-command system designed for a strict 5 GB NPU memory limit and complex multi-turn interactions.
My contribution
I worked on system deployment and memory optimization, including context compression, training/inference precision alignment, and reliable execution on the constrained NPU environment.
Highlights
- Flat-Direct single-pass agent architecture to reduce latency and cascading errors.
- Context compressed from roughly 20k to 5k tokens through semantic tool-description distillation.
- 6k-context training on an RTX 3090 with gradient checkpointing, full-linear LoRA, and BF16.
- A 20,000+ sample synthetic-data pipeline for multi-turn logic, concurrent commands, and anti-hallucination cases.
Final result: Rank 1, 100% on the local smoke test, and over 90% accuracy on complex logic cases.