Blog

Insights on edge AI testing, on-device performance, and CI/CD for Snapdragon deployments.

ProductAugust 25, 2026·7 min read

Your n8n Workflow Is Production Code. Nothing Tests It.

Edit a prompt in n8n and you have deployed to production — no review gate, no failing test, and the workflow still returns 200. EdgeGate now regression-gates any LLM workflow behind an HTTP endpoint (n8n, Zapier, Make, or anything OpenAI-compatible), with nothing to install. We changed one line of a real workflow's prompt and watched it go from "I cannot process refunds without manager approval" to "your refund has been processed" — gate RED, exit 1. Plus the finding that should change how you certify: a baseline that is already unsafe gates nothing.

Read morearrow_forward
PerspectiveAugust 14, 2026·7 min read

Your Model Matches the Reference. That Doesn't Mean It Still Works.

A new arXiv paper on cross-model KV cache transfer fit three model pairs to an identical R²=0.84 — and then retained 94%, 93%, and 37% of downstream accuracy. Across twelve evaluations, fit quality correlated with actual behaviour at r=−0.20: essentially not at all. What mattered was where the error landed, not how big it was. If you validate a quantized edge model with MSE, SQNR, or cosine similarity, that finding is about you.

Read morearrow_forward
ComplianceAugust 6, 2026·8 min read

The EU AI Act's High-Risk Deadline Just Moved. The Evidence Bar Didn't.

The EU's Digital Omnibus on AI entered into force July 27, 2026, pushing high-risk obligations from August 2026 to December 2027 (standalone systems) and August 2028 (AI in vehicles, medical devices, machinery). The reason for the delay is also the reason it doesn't buy you a pass on the evidence.

Read morearrow_forward
PerspectiveAugust 6, 2026·7 min read

NVIDIA Shrank the Robot Brain. Nobody Asked If It Still Behaves the Same.

NVIDIA's new Jetson T3000 and T2000 modules put foundation models on mass-market robots for the first time. Its own case studies show teams shrinking memory footprint to hit a lower SKU — "without compromising performance." Measured how? Compute density solves cost. It doesn't answer whether the robot still behaves the same way after the change.

Read morearrow_forward
ProductJuly 12, 2026·6 min read

Gate On-Device LLMs on Real Snapdragon — Built on Qualcomm's GenieX

Running an LLM on a phone is the easy part. Knowing it still behaves after quantization — and that it actually ran on the NPU, not silently on CPU — is what ships broken models. EdgeGate now runs a quantized LLM on real Snapdragon through Qualcomm's GenieX runtime, diffs it against a signed reference, and gates the build. Proven end-to-end on a Snapdragon X Elite.

Read morearrow_forward
PerspectiveJuly 8, 2026·6 min read

Google Benchmarked On-Device AI. Speed Was Never the Hard Part.

Google's AI Edge Portal benchmarks LLM speed on 120+ Android phones — proof the category is real. But benchmarking answers "how fast," not "did the behaviour change" after a model, firmware, quantization, or chip update. That gap, on the robots and vehicles a phone farm never sees, is the one that ships broken models.

Read morearrow_forward
ProductJuly 6, 2026·7 min read

The Black Box for Edge AI: Introducing Field Recorder

Your CI proves a model was good before it shipped. Nothing proves what it did after. Field Recorder is a black box for deployed edge AI — a tamper-evident, signed log of every on-device decision, with replay and EU AI Act Article 12 evidence.

Read morearrow_forward
ComplianceJune 23, 2026·9 min read

The EU AI Act Is Coming for Edge AI — Here's the Evidence You'll Need

High-risk AI obligations phase in through 2026–2027, with penalties up to €35M or 7% of global turnover. For on-device AI, the throughline is documented, reproducible evidence of accuracy and robustness — measured on the device, not in the cloud.

Read morearrow_forward
BenchmarksJune 23, 2026·9 min read

We Gated an On-Device LLM for Cabin Safety — Here's the Evidence

A teardown of a real Behavioral Gate run: a quantized LLM scored against a signed FP16 reference oracle on a physical Snapdragon 8 Gen 2, returning a signed, requirement-traced PASS — summary-only, no raw model output leaves the device.

Read morearrow_forward
EvidenceJune 23, 2026·8 min read

A Passing Test Is Not Evidence: The Silent Regression Problem in On-Device AI

A green checkmark proves a test ran. It does not prove what shipped to the device — which model, on which silicon, against which inputs, verifiable later. Why on-device AI needs signed, reproducible evidence.

Read morearrow_forward
ComplianceJune 23, 2026·10 min read

ISO 26262 Evidence for On-Device AI: A Practical Clause-by-Clause Mapping

A practical mapping of ISO 26262 Part 6 and Part 8 to a signed evidence bundle for on-device AI — configuration control, change management, verification with per-requirement ASIL traceability, and the honest limits of a verification tool.

Read morearrow_forward
MLOpsMay 15, 2026·9 min read

How we catch silent NPU fallback on Snapdragon in CI (and why your eval set won't)

ONNX Runtime's QNN execution provider silently routes unsupported ops to the CPU. Eval passes, production latency triples. Here's the three CI assertions that catch it before merge.

Read morearrow_forward
Founder NotesMay 4, 2026·1 min read

X - Reasoning Distillation Snapdragon On-Device Inference

1/ Reasoning distillation is quietly 2026's most important AI trend. o3-level reasoning compressing onto Snapdragon hardware. Not 2027. Now.

Read morearrow_forward
IntegrationFebruary 27, 2026·11 min read

Qualcomm AI Hub Integration Guide: From API Token to CI Pipeline

Complete walkthrough for connecting your ML workflow to Qualcomm AI Hub via EdgeGate — API setup, model submission, pipeline creation, and GitHub Actions integration.

Read morearrow_forward
QuantizationFebruary 26, 2026·9 min read

Model Quantization Testing for Edge AI: Catching Regressions Before Production

FP32 to INT8 quantization can silently break your model on real hardware. Learn how to test for accuracy drift, operator fallback, memory spikes, and latency regressions.

Read morearrow_forward
BenchmarksFebruary 25, 2026·8 min read

100 Inference Runs on Snapdragon: What the Data Shows

We ran MobileNetV2 and ResNet50 100 times each on a Samsung Galaxy S24. Here's the raw data on variance, cold starts, and why median-of-N gating catches what single-run testing misses.

Read morearrow_forward
Edge AIFebruary 22, 2026·7 min read

The Hidden Cost of Edge AI Regressions: Why Your INT8 Model Fails on Real Devices

Edge AI regressions cost teams weeks of debugging. Learn why optimized models break on real Snapdragon hardware and how automated quality gates prevent production failures.

Read morearrow_forward
TestingFebruary 21, 2026·9 min read

Deterministic Testing for Non-Deterministic Models: How Median-of-N Gating Works

Edge AI inference isn't perfectly repeatable. Learn how median-of-N gating, warmup exclusion, and flake detection bring statistical rigor to hardware testing.

Read morearrow_forward
DevOpsFebruary 20, 2026·8 min read

Evidence Bundles: Bringing Software Release Rigor to ML Model Deployment

Cryptographically signed evidence bundles give release engineers proof that every model passed quality gates on real hardware. Here's how they work.

Read morearrow_forward
Edge AIFebruary 18, 2026·6 min read

Why Cloud Benchmarks Lie About Edge Performance

Your model hits 12ms inference in the cloud. On a Snapdragon 8 Gen 3, it takes 47ms. Here's why the gap exists and what to do about it.

Read morearrow_forward
TestingFebruary 12, 2026·8 min read

Hardware-in-the-Loop Testing for AI: A Practical Guide

Emulators miss thermal throttling, firmware quirks, and quantization drift. Learn how hardware-in-the-loop testing catches failures before production.

Read morearrow_forward
CI/CDFebruary 5, 2026·10 min read

Building a CI/CD Pipeline for On-Device AI Models

Step-by-step guide to adding regression gates that test your models on real Snapdragon hardware in every pull request.

Read morearrow_forward
© 2026 EdgeGate. Powered by Qualcomm AI Hub.