embedded · ai
↗
Architected the core AI engine to translate ambiguous user inputs (mood, taste) into quantitative cocktail recipes. Leveraged prompt engineering and few-shot learning to robustly handle qualitative data uncertainty, bypassing the need for extensive model fine-tuning.
Prompt EngineeringFew-ShotEmbedded
- Problem
- Requests like "something sweet" or "I'm feeling down" have to turn into exact volumes a machine can pour.
- Approach
- The OpenAI Assistants API with a 7-shot prompt maps mood and taste to an ingredient-and-ratio JSON. An Android app handles the chat; an Arduino drives the motors and the rotating plate.
- Guardrails
- When the model's volumes added up to more than the cup holds, a post-processing step validated and renormalized the ratios. Out-of-scope requests are rejected.
- My part
- Team lead and the AI engine. We pitched it at a cocktail bar near campus and got advice on the dispensing mechanism.
- Result
- Excellence Award, 22nd Embedded Software Contest (Hanwha Vision, Nov 2024).