Dielectric

Inspiration

I've been designing PCBs for years and worked on novel ML for automating computer architecture design over the summer, but have not came across good thermal management tools for automating PCB design with AI. Also saw the YC batch 2018 company JITX and wanted to design better PCB automation software than them using modern AI and reasoning tools.

What it does

Dielectric uses computational geometry algorithms (Voronoi diagrams, Minimum Spanning Trees, Convex Hull analysis) combined with xAI reasoning to automate PCB design from natural language to production-ready KiCad files. The system analyzes component distribution, estimates optimal trace routing, detects thermal hotspots, and optimizes placement using a multi-agent architecture. Engineers can describe their PCB in plain English (e.g., "Design a multi-module audio amplifier with thermal management"), and the system automatically generates optimized layouts that are 2,000x faster than manual design.

How we built it

Backend: FastAPI with Python 3.12+, using NumPy, SciPy, and Shapely for computational geometry \ AI Integration: xAI Grok API for natural language understanding and reasoning over geometric data structures \ Optimization Engine: Simulated annealing with incremental scoring for real-time performance (<500ms for interactive optimization) \ Geometry Pipeline: Voronoi diagrams for component distribution analysis, MST for trace length estimation, Convex Hull for board utilization, and Gaussian thermal models for hotspot detection Knowledge Graph: Component relationship database that automatically identifies modules and applies design patterns \ Export: Production-ready KiCad format with proper net connections and multi-layer support

Challenges we ran into

Integrating computational geometry with AI: Feeding geometric data structures (Voronoi, MST) into xAI required careful prompt engineering to make the AI understand spatial relationships and optimization trade-offs. Scaling to large PCBs: Handling 100+ component designs required hierarchical abstraction—we implemented automatic module identification using Voronoi clustering and hierarchical optimization (modules → components). Real-time performance: Making optimization fast enough for interactive UI (<500ms) while maintaining quality required incremental scoring, caching, and a fast-path/slow-path architecture. Deterministic optimization: Ensuring same input always produces same output required seeding random number generators from user intent hashes. Visualization complexity: Plotly subplot errors when displaying multiple computational geometry visualizations simultaneously—fixed by properly structuring trace additions. Fabrication constraints: Implementing real-world PCB manufacturing limits (trace width, spacing, via sizes) and validating against industry standards.

Accomplishments that we're proud of

First computational geometry → xAI pipeline: First system to feed Voronoi diagrams, MST, and Convex Hull data directly into AI reasoning for PCB optimization. \ Multi-agent architecture: Built 5 specialized agents (IntentAgent, LocalPlacerAgent, VerifierAgent, ErrorFixerAgent, DesignGeneratorAgent) that work together like a team of engineers. \ Production-ready export: Generates KiCad files with proper net connections, multi-layer support, and manufacturing-ready output.

What we learned

Computational geometry provides structure for AI: Geometric algorithms (Voronoi, MST) create interpretable data structures that AI can reason over, bridging symbolic and neural approaches. \ Natural language + geometry = powerful: Combining human intent (natural language) with rigorous mathematical analysis (computational geometry) creates a system that's both intuitive and technically sound. \

What's next for Dielectric

Enhanced thermal modeling: Full 3D thermal simulation with ANSYS/COMSOL integration for accurate temperature prediction and heatsink optimization. \ Signal integrity constraints: Add high-speed signal analysis (impedance matching, length matching, crosstalk) to the computational geometry pipeline. \ Automated simulation integration: Direct integration with thermal and signal integrity simulators for closed-loop optimization.