Beacon

Beacon is a closed-network edge-AI system for inspecting dark environments such as factory floors and plant interiors: places where the lighting is poor, the hardware on hand is an ordinary CPU, and the video is often not allowed to leave the site. Most low-light vision today assumes a GPU and a cloud connection. Beacon assumes neither.

What it does

The pipeline has three stages, all on one machine with no internet connection anywhere. Each frame is first brightened with my training-free Retinex work: the pitch was built on XCR, and Beacon is what its stable successor, SD-Retinex, was always for. Object detection then runs on the enhanced frame. Finally, a language model on the same machine turns what the camera found into something a person on site can act on.

That last layer is still under construction. It borrows from my regional-data research: a small language model, fine-tuned on domain data and run entirely on local hardware.

Where it came from

The idea grew out of a constraint rather than a business plan. In Singapore in November 2025, a security researcher asked what I planned to do with the enhancement formula beyond publishing it. Everything I had built had to run on the device in front of me, with no cloud, and that constraint turned out to be exactly what dark, confidential sites need.

Tongali Business Plan Contest 2026

On June 20, 2026, at Chunichi Hall in Nagoya, we pitched Beacon at the final of the Tongali Business Plan Contest 2026, run by the Tongali Project (Nagoya University), as a three-person team: Uryu Den (Hiroo Gakuen), Keito Inoshita (Kansai University, my co-author on the vanishing-municipalities work), and me as team representative. The entry’s full title was “Beacon: A Proposal for a Fully Closed-Network Edge-AI System Using Low-Light Image Enhancement and a Local LLM.” We took the Tongali Award (5th place), the NICT Award, and two supporter awards, from Beyond Next Ventures and JR Central.

The Tongali contest is the Tokai-region affiliate of Kigyoka Koshien, NICT’s national student startup competition, and the NICT Award makes Beacon a candidate for the national round in March 2027. NICT decides on participation around February 2027, based on how the plan develops with the ICT mentor it assigned to the team.

Where it stands

Beacon is an early-stage student venture, and nothing is on sale yet. The work now is the language-model layer and refining the business plan with NICT’s ICT mentor ahead of that February decision.