How an AI measures in 10 seconds what takes you 3 hours

The AI stack

PlotonIQ uses specialised models for each task rather than one model for everything.

Language: Claude + Groq

Claude (Anthropic) handles all text tasks: quoting, price justification, business chat, briefing. Strong at structured output (JSON), numerical reasoning, and industry-specific instructions.

Groq serves as fallback. If Claude is unavailable, Groq with Llama takes over automatically. Slightly lower quality, but availability rises to 99.9%+.

Vision: DINO + Gemini + KI-Vision

For AI measurement, a multi-model cascade:

Claude Vision — object detection, bounding boxes.

Gemini Flash — text-prompt localisation ("find all glass surfaces").

Gemini 2.5 Flash — context analysis (window type, glazing).

KI-Vision — pixel-accurate polygon segmentation. Not rectangles — true contours. 5-15 seconds per photo.

Audio: Whisper

Whisper (OpenAI, hosted via Groq) — speech-to-text, 95%+ accuracy, all 17 languages. PlotonIQ extracts materials, dimensions, customer names and scheduling from the transcript automatically.

Production: Physics, not AI

Production scheduling deliberately uses deterministic algorithms. Machine throughput rates, setup times, deadline prioritisation — physics and logic, not pattern recognition. No API costs, no latency, predictable results.