Why
Why I built it
A missed check-in could mean forgetfulness, a phone left at home, a dead device, or someone needing help. Care staff need actionable clues without seeing full notification logs, and “insufficient information” has to be an explicit outcome rather than a forced guess.
What I built
What I built
- Care-classification pilot comparing Nemotron embeddings plus a classification head against LoRA on synthetic reporting-gap, insufficient-information, and priority-review cases.
- NV-Jev general judge adapting OpenJev-style NLI training to an NVIDIA reranker, alongside a Kev-style direct-choice pointer-head experiment.
- Evaluated a CPU service on pretrained Kev-0.8B weights to check task fit before committing to more training.
- Android minimal summaries with consent/withdrawal controls, HTTPS/SQLite backend, case triage, bounded reminders, family updates, and human sign-off.
Engineering
Problems I worked through
- Kept training, validation, calibration, and test splits strictly separate, tracking general NLI, synthetic care cases, and phone flows independently.
- The 79.7% headline score hides a gap: v1 hit only ~36% on Chinese XNLI insufficient-information cases, which matters far more for triage than the aggregate.
- On challenge cases, LoRA failed to beat a frozen backbone plus classification head, and temperature scaling hurt NLL—so the simpler baseline stayed.
- Android handles expiry and consent checks locally while model inference runs on the host for rapid weight iteration.
Role & AI
How I split the work with AI
I defined the care scenarios, privacy boundaries, and escalation rules, working with Codex across model experiments, backend, and Android. The real judgment call was separating vanity metrics from product-relevant ones and knowing when to stop fine-tuning.
Trade-offs
What testing made me change
- Only consented minimal summaries leave the phone while raw personal data stays local; missing evidence stays explicit for human follow-up.
- Host benchmarks run ~60 ms per batch on v1 (RTX 4060) and 0.6–1.4 s on CPU with Kev-0.8B, letting me validate judgment logic and workflow first before tuning mobile latency.
Outcome
Outcome
Delivered trained weights, item-level evaluations, a CPU inference service, and end-to-end workbench flows on synthetic data; next up is moving inference on-device and validating in real care settings.