AI research · 2026
Ganpati blessing booth
82 captures, 73 portraits
The problem
Putting a generative AI experience in front of eighty guests, in a house, on one laptop, with a deadline that could not move.
The system
An ESP32 camera over wifi, everything running locally with no cloud service anywhere in the pipeline, and a two-stage reveal: a plain blessing card within seconds of the namaskar, quietly replaced by the AI portrait when it was ready.
What happened
82 captures, 73 portraits. The four that could not be recovered were framing failures, not model failures: heads at the top edge of the frame, a body running off the left side, a camera aimed slightly at the ceiling.
Not claiming
One pipeline, one similarity threshold, one evening. The observation that styles changing a guest’s clothes lose their likeness rests on three attempts and is a hypothesis, not a finding.