Start in your own words
A fine-tuned NVIDIA 4B model scores eight intents in one forward pass. Below 0.85, NVIDIA 30B on Nebius provides fallback. If reliable routing is unavailable, ask or hand off.
約好約HAOYUE
Understand the correction. Protect the appointment.
A dental concierge for Taiwan · Live conversation, visible decisions
Try: “Not cancel—please reschedule”
Comparison & limitations ↓Choose a patient → send a request → pick a time → review and confirm. Watch when the saved booking actually changes.
HAOYUE Dental
Fictional clinic · Demo accountSimulated LINE · Synthetic patients · No real bookings
Whose day can we make easier?
Connecting and reading appointments…
Free text works too. Try a scenario:Your demo bookings are independent and survive reload. Switching patient starts a new demo. Please avoid real personal data.
SMALL DECISIONS. REAL CONSEQUENCES.
A good reply is only half the job.
“Not cancel—please change the time.” People correct themselves, worry about visiting, and sometimes click by mistake. HAOYUE connects intent, patient guidance and verifiable booking transactions.
Be patient without profiling people
Explicit worry receives reassurance. Confusion gets one-step guidance. Repeated requests skip another model call. Style cues use rules, not a trained emotion classifier or an age guess.
Confirm, commit, then read back
Application code checks ownership, availability and versions. Confirmation precedes an atomic Supabase update. Conflicts, replays and stale confirmations are tested. The panel shows actual events.
Does the specialised model add value?
The same 64 fixed synthetic inputs and labels. Rules are executed again; NVIDIA predictions come from the saved CUDA experiment. This compares single-message intent, not complete booking success.
Rules
NVIDIA 4B
NVIDIA 4B + LoRA
| Model | Accepted | Wrong among accepted |
|---|---|---|
| NVIDIA 4B | 30 / 64 | 2 |
| NVIDIA 4B + LoRA | 54 / 64 | 1 |
One high-score error remains. Every write still requires validation and confirmation. These author-written cases are not independent clinical validation. Model rows omit the app’s safety overrides; rules include handwritten safety routes. Equal bilingual quality, workload reduction and cloud cost saving remain unmeasured.
Inspect inputs, labels and saved predictions ↗What we test when things go wrong
- Conflicting slots and stale confirmations cannot overwrite the current booking.
- An identical confirmation is replayed without another write.
- Unclear intent asks for clarification. Symptoms and family requests go to a human.
- Unavailable models and expired budget checks do not become invented successful decisions.
- Visitor workspaces are independent; same-workspace conflicts remain protected.
58 / 64v2 synthetic holdout; base 52/64
53 msMedian local Mac inference; excludes network
127 sL4 v2 training and evaluation segment
397,312Trainable LoRA parameters
Stack: NVIDIA Nemotron Nano 4B (LoRA / MLX) → NVIDIA Nano 30B fallback on Nebius Token Factory → confirmation and atomic transactions → Supabase. Vercel and Supabase run in Tokyo; authenticated Cloudflare Tunnel connects Mac inference.
Tavily powers staff research: sources become drafts and require human approval before entering the knowledge base. Patient messages do not trigger web search. Shared budgets and source/session limits protect demo credits.
Evidence, honest limits and reproducibility
The model was tuned primarily on synthetic Taiwanese Mandarin. Equivalent bilingual model performance is not established. The 64 author-written synthetic items are not clinical validation. One of 54 accepted high-score predictions was wrong, so changes still require confirmation.
Inspired by Kev/Jev, this implementation projects the final hidden state onto eight choices. It does not implement Kev’s masked multi-query architecture. The first version’s overconfidence failure and the second experiment are documented.
The Quick Tunnel has no production availability guarantee. Real deployment still needs patient authentication, clinic review and stable model hosting.
Source, training recipe and reports ↗ · Download the adapter ↗
Read measured evidence JSON ↗