A generative recipe engine that speaks motor RPM.
How we built upliance AI — a hardware-constrained personalization engine that turns natural-language cooking requests into safe, machine-executable workflows for the upliance smart appliance.
- CLIENT
- upliance
- INDUSTRY
- Smart kitchen hardware
- MODEL
- Self-hosted Qwen on vLLM
- STACK
- pgvector · Zod · WebSockets
{
"step": 4,
"heat": { "stage": "saute", "temp_c": 160 },
"stir": { "rpm": 40, "reverse": true },
"scale": { "prompt": "tofu", "g": 400 },
"safety": { "dry_run_lock": true }
}Free text can’t drive a heating coil.
The upliance appliance runs on precision actuators — automated stirring blades, induction heating and integrated weighing sensors. But it relied on static, hardcoded recipe programs that couldn’t handle real kitchens: keto, vegan or Jain diets, missing pantry items, portion changes or personal spice tolerance.
Generic chat LLMs weren’t the answer. Unconstrained text generation can’t be safely translated into the strict thermal limits, blade speeds and timed physical steps the hardware needs.
Four layers between a craving and a motor.
Semantic recipe discovery
pgvector RAG over the verified upliance recipe base, matching dietary profile, pantry and time constraints to the best base program.
Context-aware adaptation
A self-hosted Qwen on vLLM recalculates ingredient proportions, liquid-to-solid ratios and cooking duration on the fly.
Deterministic instruction compiler
Strict Zod schemas turn steps into motor speeds, multi-stage heat curves, timers and scale/tare prompts.
Real-time appliance sync
Lightweight WebSockets to the touchscreen OS for live guidance, appliance control and taste-adjustment feedback.
From natural language to machine code — with a safety envelope around everything.
- LAYER 1User intentTouchscreen / voice
- LAYER 2ASemantic RAGPostgres + pgvector HNSW
- LAYER 2BAdaptation engineQwen via vLLM
- LAYER 2CHardware compilerZod enforcement
- HARDWAREupliance applianceActuators & sensors
| LAYER | SUBSYSTEM | CORE TECHNOLOGY | RESPONSIBILITY |
|---|---|---|---|
| 1 | User intent ingestion | Touchscreen / voice NLP | Substitutions, dietary preferences, portion scaling |
| 2A | Semantic discovery & RAG | PostgreSQL + pgvector HNSW | Matches taste & pantry to certified base recipes |
| 2B | Culinary adaptation engine | Self-hosted Qwen via vLLM | Recalibrates volume, moisture, duration & thermal curves |
| 2C | Deterministic hardware compiler | Zod schema enforcement | Validated machine code for RPM, heat curves, scale calibration |
| SAFE | Hardware guardrail engine | Embedded firmware interlocks | Burn, boil-over, dry-run and lid-lock protection |