cd /news/machine-learning/seeking-cs-lg-endorsement-on-device-… · home topics machine-learning article
[ARTICLE · art-134888] src=discuss.huggingface.co ↗ pub= topic=machine-learning verified=true sentiment=· neutral

Seeking cs.LG endorsement: on-device glucose trend forecasting and INT8 quantization

Independent researcher Mohamed Menasy is seeking a cs.LG endorser on arXiv for "GlucoEdge," an engineering study of on-device five-class glucose trend forecasting from continuous glucose monitor data using a 2,909-parameter 1D CNN. Static INT8 post-training quantization cut the model artifact size by about 31% but dropped macro recall from 0.5060 to 0.4135, with recall falling across all four directional classes. The work covers PyTorch-to-LiteRT conversion, calibration-data coverage and input-range clipping, cross-runtime numerical parity testing, and offline inference on a physical Android device, with code and reproducibility materials on GitHub under endorsement code PC9MML.

read1 min views1 publishedSep 20, 2026

Hi everyone,

I’m an independent researcher preparing my first arXiv submission in cs.LG, and I’m looking for an eligible endorser willing to look at the work.

Paper: GlucoEdge: An Engineering Study of On-Device Five-Class Glucose Trend Forecasting and INT8 Quantization

It covers an end-to-end on-device ML pipeline for glucose trend forecasting from continuous glucose monitor data: five-class, 15-minute trend forecasting from CGM history; a compact 1D CNN with 2,909 parameters; PyTorch to LiteRT conversion; static INT8 post-training quantization; the effect of quantization on minority-class recall; calibration-data coverage and input-range clipping; cross-runtime numerical parity testing; and offline inference on a physical Android device.

INT8 quantization cut the model artifact size by about 31%, but macro recall dropped from 0.5060 to 0.4135, with recall falling across all four directional classes. So the study looks at deployment size alongside what quantization does to minority-class behavior.

I’ve tried to be explicit about the limitations: overlapping time-series windows, calibration-set selection, context-prefix leakage risk, single-device latency measurements, and the fact that this is a research prototype rather than a medical system.

Code: GitHub - mohamedmenasy/GlucoEdge · GitHub (Android implementation, training and conversion code, tests, model artifacts, and reproducibility materials) Requested category: cs.LG

Endorsement code: PC9MML

If you’re eligible to endorse in cs.LG and willing to look at the manuscript, I’d appreciate it. I’m also glad to get technical feedback even if you can’t endorse.

Thanks,
── more in #machine-learning 4 stories · sorted by recency
── more on @mohamed menasy 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
Live at https://your-agent.zahid.host
Get free account → Pricing
from €0/mo · no card required
LIVE [news/seeking-cs-lg-endors…] indexed:0 read:1min 2026-09-20 ·