darmousseh.com — Engineering Blog & Live Demos

Engineering Blog & Live Demos

Fast local ML, calibrated embeddings, and small sharp systems.

Technical writeups with real benchmarks and working edge demos you can test right in the browser.

Featured PostAll posts →

Post #1 2026-11-21 · 10 min read · systemone · EmbeddingGemma 2

System One: Hill-Climbing Multimodal Embeddings from 77% Zero-Shot to 99% Calibrated Classification in 30 μs

How a frozen multimodal encoder (google/embeddinggemma-2) paired with calibrated linear probes, Wilson statistical gates, and leave-one-group-out novelty floors climbs from 77.4% zero-shot to 99.0% Top-1 (100% answered accuracy, 100% OOD vegetable rejection) in a 241 KB binary file running in 29.8 μs on CPU. Includes the modality-gap collapse trap, few-shot curves, and deploying natively on Cloudflare Workers AI.

77.4% → 99.0%Fruits 360 25-class Top-1
100.0%Gate@0.90 accuracy (947/947)
29.8 μsSingle-image CPU readout
100%OOD rejection (0/100 false)

Interactive Demo · System OneAll demos →

Live Fruit Classifier & Novelty Gate

Runs @cf/google/embeddinggemma-2 on Cloudflare Workers AI and executes three binary .system1 readouts (~28 μs) on the same 768-d vector.

Live Edge API
Tap a sample or upload a photo Test zero-shot traps (grapefruit, grape, mango) and OOD vegetables (onion, potato).
Click any sample to classify live10 curated cases
Want interactive threshold & novelty sliders, session history, and all 25 classes? Open full interactive demo →

Recent PostsArchive →

System One: Hill-Climbing Multimodal Embeddings to 99% in 30 µs

How frozen EG2 multimodal embeddings + calibrated linear probes, Wilson gates, and novelty floors climb from 77.4% zero-shot to 99.0% Top-1 (100% OOD rejection) in a 241 KB artifact.

2026-11-21

DemosAll demos →

Fruit Classifier

Snap or upload a fruit photo for instant identification across 25 common fruits, with automatic rejection of unknown items.

2026-11-21