HaLowLink ViewSense
A complete AI-powered visual sensing platform. Ultra-low-power sensors on Wi-Fi HaLow push frames through a real-time MQTT broker into parallel AI analysis pipelines, triggering smart actions instantly.
The first generation of IoT was blind — temperature, humidity, motion. The next generation sees. ViewSense gives GenAI its eyes, turning every visual sensor into an intelligent observer that understands the world in real time.
Temperature, humidity, motion.
Simple scalars. No context.
Frames over HaLow. Rich
visual data at ultra-low power.
AI understands every frame.
Counts, reads, detects, acts.
Until now, deploying visual sensors at scale was simply not possible. Cameras meant mains power, structured cabling, costly installation, and limited coverage — restricting monitoring to a handful of fixed points. Wi-Fi HaLow (IEEE 802.11ah) removes every barrier. For the first time, compact battery-operated visual sensors can be placed anywhere — no wires, no electricians, no infrastructure — blanketing entire facilities with intelligent coverage and giving Agentic AI complete visibility of the physical world.
Previously, every camera required power and ethernet — limiting placement to wherever wiring could reach. HaLow's extreme power efficiency means visual sensors run for months on AA batteries. Mount one in minutes, not days. No cables, no contractors, no compromise on placement.
Legacy systems meant a few expensive cameras covering a fraction of your space. HaLow supports thousands of devices per access point, with sub-1 GHz signals that cut through concrete, steel, and vegetation. One AP blankets an entire warehouse, campus, or farm — eliminating blind spots entirely.
Traditional visual monitoring was expensive — large devices, professional installation, ongoing power costs, limited scalability. HaLow enables small, low-cost sensors with minimal batteries, no cabling, and self-install deployment. The result: 10x more coverage points at a fraction of the cost, making facility-wide Agentic AI monitoring economically viable for the first time.
Every dimension that matters for battery-powered visual IoT at facility scale.
| Dimension | Wi-Fi HaLow | BLE 5.x | Zigbee | Z-Wave | LoRaWAN | LTE-M | 5G RedCap | Satellite |
|---|---|---|---|---|---|---|---|---|
| Range (outdoor) | 1–3 km | 200–400 m | 100–300 m | 100 m | 5–15 km | 10–15 km | 5–10 km | LEO footprint |
| Indoor / walls | 300–500 m, 4+ walls | 30–50 m, 1–2 walls | 10–30 m, 1–2 walls | 30–50 m, 1–2 walls | 100–200 m, 2–3 walls | 200+ m, 3–4 walls | 150–300 m | N/A (sky view) |
| Max nodes / AP | 8,192 | 7 (mesh ~100) | ~250 | 232 | ~50,000 | ~100K/cell | ~100K/cell | 100s/beam |
| Bandwidth | 43 Mbps (4 MHz) | 2 Mbps (theoretical) | 250 kbps | 100 kbps | 50 kbps (SF7) | 1 Mbps | 150 Mbps | Variable |
| Image transfer | Yes (~40ms) | Marginal (mins) | No | No | No | Yes (2–3s) | Yes (<0.1s) | Weather dependent |
| Battery life | 2–5 years | 6–12 months | N/A | N/A | N/A | 3–6 months | 1–3 months | Days–weeks |
| Recurring cost | $0 | $0 | $0 | $0 (proprietary) | $0–5/mo | $1–10/mo | $5–20/mo | $5–100/mo |
| Security | WPA3-SAE | AES-128-CCM | AES-128 (CVEs) | S2 AES-128 | AES-128 OTAA | LTE SIM-based | 5G-AKA 256-bit | Proprietary, varies |
| OTA updates | Yes (MBs in seconds) | Slow (mins/100KB) | Impractical | Very slow | No (tiny payload) | Yes | Yes | Unreliable |
| Always addressable | Yes (TWT) | No (per-session) | Partial (polling) | Yes (power cost) | No (uplink only) | Partial (eDRX) | Yes | No (intermittent) |
| IP native | Yes | No (GATT gateway) | No (gateway) | No (proprietary) | No (server) | Yes | Yes | Varies |
Visual IoT demands image-grade bandwidth AND multi-year battery life AND facility-wide range through walls AND thousands of nodes per access point AND zero recurring fees. Every alternative fails on at least one axis: LoRa/Zigbee/Z-Wave can't carry images. Cellular imposes per-SIM fees that destroy unit economics at scale (1,000 cameras × $5/mo = $60,000/year). Bluetooth collapses beyond 50 m indoors. Satellite requires sky view and costs a fortune. Wi-Fi HaLow is not merely better — it is the only technology where the physics, economics, and protocol architecture simultaneously satisfy the requirements of battery-powered, facility-scale, GenAI-powered visual monitoring.
Every frame travels from visual sensor to intelligent response in a fully automated, fault-tolerant pipeline.
Battery-powered visual sensors with 6+ month life and colour night vision. Intermittent JPEG frames, not video — maximising battery and bandwidth.
Wi-Fi HaLow (802.11ah) sub-1 GHz wireless penetrates 4+ walls and reaches 1 km. Frames arrive via MQTT QoS 1 through an embedded amqtt broker.
Frames hit AWS Bedrock in parallel output-type groups (count, reading, status, detection). Each group gets a specialised Chain-of-Thought prompt. Results are whitelist-filtered and readings are median-stabilised.
Results persist to SQLite (WAL mode) with in-memory caching. SSE pushes frame, AI, and activity events to connected dashboards in real time — no polling.
Configurable action rules fire on detection — direct-address voice deterrence, sirens, lights, relay switches, push/SMS/email notifications, webhooks, and activity logging.
From precision instrument reading to full-text scene search, every capability is production-hardened and built on real AI engineering.
Detectors are grouped by output type — count, reading, status, detection. Each group fires a dedicated Bedrock call simultaneously. Latency equals the slowest group, not the sum.
Instrument readings use a structured Chain-of-Thought: identify type, enumerate all scales, select inner scale, map markings, locate needle, calculate with arithmetic, sanity check. Dual-scale gauges read correctly every time.
Every frame gets a parallel Bedrock call producing a scene caption and object list — indexed by SQLite FTS5 for instant full-text search. Find "red hat near loading dock" across thousands of frames.
Configurable trigger-action rules with zone, sensor, and detector filtering, 3 schedule modes (always, active hours, after hours), escalation chains, and cooldown windows.
Server-Sent Events replace all polling. Frame, AI, and activity events push instantly to connected browsers. A ring buffer enables reconnection replay via Last-Event-ID.
A guided wizard with canvas-based floor plan editing, drag-and-drop device placement, industry-filtered detector selection, live cost estimation, and one-click activation.
A per-site switch that makes frames analysis-only: AI extracts the intelligence in-memory, then the raw image is never written to disk and never exposed via API, SSE, MCP, or search. Store the insight, not the pixels.
The platform watches its own health and dependencies. A live model probe catches a deprecated AI model before it silently breaks analysis; a disk guardian sheds disposable data under pressure so the server never falls over; a threshold watchdog flags any queue, pool, or error rate approaching its limit.
A 26-tool Model Context Protocol server lets Claude and ChatGPT see camera feeds, query AI history, and reconfigure sites by voice or chat. Dual transport (SSE + StreamableHTTP), OAuth 2.1 for ChatGPT, and per-site role-based access — one identity across every interface.
Every layer chosen for reliability, performance, and zero operational overhead.
Live platform statistics loaded from the running system.
ViewSense isn't a prototype — it's a fully engineered platform with comprehensive test coverage, hardened security, and real-time observability. Every line of code is tested, every endpoint is authenticated, every action is auditable.
From detector selection to action rules, everything is configurable through the UI or natural language via AI assistants. Deploy a complete monitoring solution without writing a single line of code — the 15-step wizard handles everything.
Whether monitoring a single loading dock or a multi-building campus, the same platform scales effortlessly. Industry-specific detector presets and templates mean you're operational in minutes, not months.
Visual sensing delivers the insights of continuous surveillance with none of the privacy burden. Frames are analysed by AI and discarded — no recordings, no video streams, no raw footage stored beyond a configurable retention window.
Every API request is verified against session tokens with configurable expiry. Auth middleware enforces access on 85+ endpoints — only public reference data is unauthenticated.
Every response carries Content-Security-Policy, X-Frame-Options, X-Content-Type-Options, and Referrer-Policy headers — mitigating XSS, clickjacking, and MIME sniffing attacks.
Login and registration endpoints enforce per-IP rate limits — 10 attempts per 15 minutes. Brute-force attacks are stopped before they reach the authentication layer.
All JSON writes go to a temporary file then atomically replace the target via os.replace(). SQLite runs in WAL mode. No partial writes, no corruption — even under sudden power loss.
All query parameters pass through safe_int() with enforced bounds. Raw int() is never used on user input — eliminating injection and overflow vectors at the boundary.
Saved frames are automatically purged after a configurable retention window (default 7 days). No indefinite image storage — the system retains AI insights, not raw visual data.
Launch the app, create a site from a template, and watch the AI pipeline light up in real time.