Inspiration The "hallucination" of LLMs isn't just about text; it's about the physical world. While building an automated robot arm, I noticed that most models couldn't distinguish between a toy-grade servo and an industrial-grade motor. They might suggest a $5 hobby servo to lift a 1kg load at a 1m distance—a design guaranteed to end in smoke and fire. I was inspired to create PhysiLink, a safety layer powered by the Gemini 3 API that forces AI to respect the laws of physics before any hardware is purchased.

How I built it PhysiLink is built on a "Reasoning-to-Verification" pipeline:

Extraction Layer: Using the Gemini 3 API, I developed a sophisticated system instruction set that deconstructs vague user prompts into structured physical parameters like mass, length, and velocity.

Physics Engine (Code Execution): Instead of letting the LLM "guess" the result, I utilized Gemini's built-in Code Execution capability to run real-world Python scripts. It calculates required torque based on lever arm physics and applies a mandatory 3.0x safety factor.

Grade Classification: The system maps calculated values against a hardware threshold database to categorize designs as "Toy," "Prosumer," or "Industrial."

Challenges I faced The biggest challenge was Semantic Mapping. Mapping subjective words like "fast" or "heavy" into specific SI units (m/s or kg) required intense prompt engineering and few-shot examples. I had to ensure the Gemini 3 API makes conservative, safe engineering assumptions rather than the optimistic "hallucinations" typical of standard LLMs.

What I learned I learned that the true power of the Gemini 3 API isn't just in its chat capabilities, but in its ability to act as a bridge between unstructured intent and structured physical constraints. Combining LLM reasoning with deterministic code execution is the only way to achieve true "Physical Grounding."

Built With

Share this project:

Updates