Private WebML Demo
Classify photos and chat with LLMs running entirely in your browser tab.
Try live inference →
Intelligence that lives at the edge.
EdgeBites designs, deploys, and operates AI where your data is born — on-device inference, microcontroller integration, and production edge software. No cloud round-trip required.
Edge NEWS
Machine-filtered edge-AI, MCU, and NPU headlines — hourly.
Read the news →
Micro-Blog
Short notes from the lab — deployments, measurements, things learned the hard way.
Read the blog →
Edge Solutions
Architecture, deployment, optimization, microcontrollers, and security — edge AI from sensor to insight.
Overview · Deployment · Optimization · Microcontrollers · Security
Prefer a reader? Micro-Blog RSS feed · Full reading setup at rss.edgebites.com · Fediverse: social.edgebites.com
Edge Solutions Expertise
Edge AI systems for places the cloud can't reach — factory floors, farm fields, wearables, vehicles, and remote sites. We take you from sensor to insight.
Architecture & Sizing
Start with the BOM, not the model. Hardware selection, latency and power budgets, costed before silicon.
Model Optimization
INT8 quantization, pruning, distillation, NPU delegation — every accuracy point accounted for.
Fleet Operations
OTA with health gates, staged rollouts, automatic rollback. Proven on 1GB hosts — like this one.
Security by Default
Secure boot, encrypted artifacts, mutual TLS. Designed in, not patched on.
Edge AI Software Deployment
A shippable path from notebook to device — the same pipeline discipline we run ourselves.
- Profile — benchmark candidate models on target silicon, never on a workstation.
- Optimize — quantize and delegate operators to NPUs and DSPs; verify the accuracy delta.
- Package — reproducible builds, signed artifacts, and SBOMs for every release.
- Roll out — staged over-the-air delivery with health gates and automatic rollback.
- Observe — edge telemetry, drift alerts, and retraining triggers that close the loop.
This site is served from a 1GB Debian host using exactly this discipline. Ask for a deployment review →
Complete Optimization
Precise optimizations at every level — from sensor hardware to edge terminals to cloud servers. No watt or millisecond wasted.
Sensor Hardware
Spend microwatts, not milliwatts: adaptive sampling, duty cycling, wake-on-event.
Firmware & TinyML
Shrink models 4× with INT8, prune dead weights, delegate to NPU/DSP/CMSIS-NN.
Edge Terminals & Gateways
Cache, batch, forward: MQTT/CoAP/LoRaWAN tuned per link budget.
Edge Servers
Batch inference, pack accelerators, degrade gracefully when nodes drop.
Cloud Servers
Autoscale to zero, cache artifacts, burn preemptible watts for batch.
Measure Everything
Power traces, latency budgets, accuracy deltas — unmeasured is unoptimized.
Tokens per Watt (TPW) Optimizations
The efficiency metric for the LLM era: useful tokens per joule, from battery-powered MCUs and sensors to edge terminals, cloud servers, and GPUs. Right-size the model, quantize aggressively, batch where latency allows — and put each token on the cheapest watt that meets the deadline.
Private WebML — Live In-Browser Inference
Real models running 100% on your device — vision and text inference with nothing uploaded. Pixels and prompts never leave this browser.
Device check
Probing…
Description appears here after you load the VLM and pick an image…
Description:
Drop an image here or use the picker above
Community results load here for the selected model + context.
AI Search
Browse Google AI Mode, then push the page here with the bridge icon (or Ctrl+Shift+E). The local WASM filter renders text windows + prompt and removes side panels — everything stays in this window, nothing leaves your browser.
New thread — open Google AI Mode, then push the page here with the bridge icon (or Ctrl+Shift+E).
Blocked: scripts, trackers, nav, side panels, ads — only text windows + live prompts. Nothing leaves your browser.
Drop a saved .html file here — parsed 100% locally by WASM
Bridge pull setup (fallback — push needs no ID)
Fallback path only. Normal flow needs no ID: install edgebites-ai-bridge.zip (unzip → chrome://extensions → Developer mode → Load unpacked), open a Google AI Mode tab, click the bridge icon or press Ctrl+Shift+E.
Push flow: install edgebites-ai-bridge.zip, open Google AI Mode, click the bridge icon (or Ctrl+Shift+E) — the page renders here, panels removed. Direct page fetch is blocked by Google CORS, so the companion extension carries the HTML inside your own browser. This page never opens new windows itself. Extractor: /assets/ai-extract/pkg/ai_extract_bg.wasm.
Microcontroller Integration Projects
Milliwatt-scale intelligence — sensing, fusion, and control on microcontrollers.
TinyML Sensing
Vibration, acoustic, and vision anomaly detection that sips power on Cortex-M and ESP32-S3.
Sensor Fusion
IMU, environmental, and GNSS fused on-device; only events use the uplink.
Edge Gateways
Buffer, store-and-forward, translate: MCU-to-Linux plumbing.
Security and Traffic Control
Where it matters most, at the edge — protection and prioritization computed on-device, not in a distant cloud.
Zero-Trust Device Identity
Every node carries keys and attests firmware. No trust, not even on LAN.
Encrypted Traffic & mTLS
Mutually authenticated, encrypted transit — our own mail and admin ports run TLS-only.
Traffic Shaping & QoS
Alerts jump the queue; telemetry waits. Lossy links stay useful.
Secure OTA & Rollback
Signed artifacts, staged rollout, auto-revert on trouble.
On-Device Detection
Local anomaly models flag intrusions in milliseconds — offline included.
DDoS Absorption
Rate limits and edge filtering keep fleets answering when the internet won't.
Open Source Projects
We build in the open — tooling, examples, and building blocks under open licenses. Start here:
github.com/EdgeBites — all repositories →
Edge Deployment Tooling
Reproducible single-host edge stacks — the configs serving this page.
TinyML Examples
Reference MCU integrations published with measured latency and power.
Self-Hosted Building Blocks
RSS, federation, and mail glue for a presence you fully own.
Micro-Blog
Short notes from the lab — deployments, measurements, and things we learned the hard way. Also available as RSS.
Loading posts…
NEWS
Edge-AI, MCU, and NPU headlines from across the web — machine-filtered for signal, refreshed hourly. Full reading in FreshRSS.
Loading headlines…
Download OPML to import this feed list into any reader.
Contact
Tell us about your edge project — what senses, what decides, what constrains it. We reply by email, usually within two business days.
Prefer the fediverse? social.edgebites.com