How the Insane Speed of AI Development Has Changed an AI Company

July 5, 2023

Approximately 70% of daily work at Fyma now involves AI tools - a figure that would have been unimaginable eighteen months ago. The speed of generative AI development has shifted the landscape of what previously felt like insurmountable challenges, particularly in data collection and model training for computer vision applications.

ObjectX: The Secret Project

Fyma developed an internal initiative to enable customers to detect any object using only 10 uploaded images. The workflow: automatic labelling of provided images, generation of approximately 2,000 synthetic variations via AI, automated model training for object detection, and immediate deployment through Fyma's computer vision pipeline. Customers now only need to connect cameras, upload reference images, and define automation rules.

Testing with Real Objects

Initial testing used a yogurt bottle as proof-of-concept - an intentionally mundane object with no existing training data. Results demonstrated that AI-generated training datasets could effectively teach detection models. The synthetic data generalises the model better than traditional datasets built from painstakingly collected real-world images, because it systematically covers edge cases that manual collection would never encounter.

Production Deployment

Recent applications include airport equipment detection - an environment where privacy concerns make real-world image collection particularly difficult. The system successfully combines AI-generated, synthetic 3D, and real-world imagery automatically. What previously required months of fieldwork and expensive annotation now takes 48 hours from brief to deployment.

What This Means for Computer Vision

The ability to train accurate custom detectors from minimal seed images fundamentally changes the economics of computer vision deployment. Projects that were previously uneconomical - too narrow in scope to justify the data collection cost - are now viable. This democratises access to precise, purpose-built analytics for organisations without large ML engineering teams.

The Broader Implication

AI development is compounding. Each capability unlocked - synthetic data generation, automated labelling, rapid model iteration - makes the next capability easier to build. For Fyma and the operators who use the platform, this means the question is no longer 'can we detect this?' but 'what should we be measuring that we haven't thought to measure yet?'