05 · 2026SoftwareAI
Laundry Manager (LoadCheck)
Deterministic Bilingual Voice & Offline-First Garment Load Engine
Role & Context
Full-Stack & NLU Systems Architect
Verified Local Build (Next.js 16.3 + Dexie + Qwen3-ASR / LiveKit) · Private Repo
NLU Determinism
Rule-Audited (Zero LLM State Edits)
Speech ASR
Optional Qwen/Qwen3-ASR-1.7B Local WS
Storage
Dexie IndexedDB (No Cloud Account)
Repo Visibility
Private Repository · Inspected Locally
Bilingual English + Hinglish NLU Sandbox
Pipeline Output Trace
Ready · Deterministic pipeline mirrors loadcheck IntentDetector.ts
Simulated Drum Load4.05 / 7.00 kg (58%)
Cotton Shirts4 × 0.25 kg
Denim Jeans2 × 0.65 kg
Bath Towels2 × 0.45 kg
Bedsheets1 × 0.85 kg
Problem & Engineering Context
Generic LLM tool calls hallucinate quantities or concatenate numbers across speech chunks ("6" + "6" -> 66) or misread Hinglish household requests ("aur ek bedsheet daal do", "make machine capacity 7 kg"). LoadCheck enforces deterministic parsing before mutating IndexedDB.
Architecture & Key Implementation Details
- Decoupled pipeline: both Voice (`Qwen3-ASR 1.7B` over WebSocket / optional LiveKit) and in-browser Chat emit `NormalizedUserInput` into the identical TypeScript NLU stack.
- Stage sequence: TextNormalizer -> IntentDetector -> NumberResolver -> ReferenceResolver -> OperationPlanner -> SemanticSafetyGate -> TransactionManager.
- Verified 21 unit test suites covering number safety, pronoun resolution, noise robustness, multi-item transactions, and reversible one-click undo. Note: source repository (`laundry-Model`) is private.
Key Deliverables
- English + Hinglish Intent Parser (`IntentDetector.ts`)
- Drum Capacity & Dry-Weight Balance Engine
- Atomic Undo & Clarification Banner State Machine
- Interactive Deterministic NLU Sandbox Embedded Below
Tools & Stack
Next.js 16.3React 19TypeScript 5Dexie (IndexedDB)Qwen3-ASR 1.7B / LiveKitVitest (21 Test Suites)