Intelli-Gen

My first startup, funded under Gujarat’s Student Startup and Innovation Policy. It turns 5–15 reference photos into a full training dataset for machine vision, and the research behind it became a paper.

Founder and lead author. Ran the research across the generative image stack, built the fine-tuning and generation pipeline, and led the same work earlier as an intern at SPAN Inspection Systems.

The problem

Inspection models need thousands of pictures of defects. Factories have a few, because defects are rare and stopping a line to photograph them is expensive.

Four months of research first

Sixteen to eighteen weeks went into research alone, working through the generative image stack: GANs, VAEs and diffusion models, and three ways of fine-tuning them, DreamBooth, LoRA and Textual Inversion.

Teach the model the part

A diffusion model is fine-tuned on a handful of real photos until it has learned that specific product: tablet shapes, capsules, foil types.

Ask for the defects

Each defect class gets its own prompt: missing pills, broken tablets, colour variation, impurities. Every generated image arrives already labelled.

Hand over a dataset

The study produced more than 2100 unique images, varied in colour, shape and packaging, and reviewed by inspection engineers at SPAN Inspection Systems.

Written up as a paper

“Transforming Dataset Creation: Leveraging Generative AI for Machine Vision Applications in Data-Limited Domains”, with Kandarp Joshi, Juhi Gajjar, Riddhi Parmar and Pranay Soni of SPAN Inspection Systems.