Oct 2026

GreenMind AI Hackathon 2026

Cut the energy per answer of a 9B Thai and English LLM by 66% on one NVIDIA A100 GPU.

Team: Team FC P'TAE (Team 18)
Result: Scored 96/100
Links: GitHub Model on Hugging Face

Ronnapee presenting for Team FC P'TAE on stage at the GreenMind AI Hackathon, with the slide “One change at a time, measured”

Overview

What it is

A two-day hackathon hosted by CMKL University with ExxonMobil Thailand and the AIEI network. The task: serve a rush of 300 requests in 2½ minutes on one NVIDIA A100 with fewer joules, on time, without making answers worse.

Highlights

What we shipped

Scored 96/100
Energy per answer down from 467 J to 157 J, with on-time answers up from 64% to 98%
Speculative decoding, chunked prompt reading, a thinking budget and our own INT8 W8A8 weights on stock vLLM
Accuracy held at 88.3% against 87.2% for the default server
46 experiments and 9 official runs, each measured on the GPU's power meter

Gallery

Photos

Ronnapee (right) with Team FC P'TAE in front of the GreenMind banner with the ExxonMobil, CMKL and AIEI logos — Team FC P'TAE at GreenMind AI Hackathon 2026

Ronnapee opening the pitch in front of the title slide — Opening our final pitch

Ronnapee explaining the one-GPU setup during the pitch — One NVIDIA A100, 300 requests in 2½ minutes

Results slide: official grader score rising from 0 with the default server to 31.5 with guessing ahead, 59 with chunked prompt reading, 88.0 with a thinking cap and 99.40 with 8-bit weights, as energy per answer falls from 467 J to 157 J — Four changes took energy per answer from 467 J to 157 J

The Newton VEX team on stage with their robot and award in front of the RAT Foundation screen

Next project

VEX V5 Robotics Competition

Designed and programmed the team's robot for a national-level metal robotics competition.

Contact

Let's build something.

Email me

Bangkok, Thailand