On-prem · Edge AI · No cloud required
r450k is a suite of apps that install on hardware you already own: AI cameras, building sensors and badge readers at the edge, one control plane that adopts them all and correlates events across them, and single sign-on tying it together. It scales from a single Raspberry Pi watching a couple of cameras to a GPU server running real-time AI across dozens. Your footage — and everything else — never leaves your network.
Entry tier runs on a Raspberry Pi 4 (a couple of cameras with motion or nano-model AI). Real-time, multi-camera detection and teaching the camera your own skills scale up to mini-PCs and NVIDIA GPUs. The sensor hub, door controller, identity broker and control plane are single pure-Go binaries and need far less.
Capabilities
Cameras, sensors, doors, identity and fleet management — built to run on small devices, entirely on your own network.
Two-axis rules pair a mode — presence, crowd, intrusion, line crossing, multi-line crossing, or LPR — with target classes from a data-driven registry, scoped to multiple detection zones per rule. Fire and smoke ship as first-class event classes, and an "Anything" wildcard fires on any detected object.
A second-stage plate detector + OCR on a dedicated high-res frame fires on any readable plate or only a watchlist. Fuzzy matching absorbs OCR errors; plate text, vehicle type, and color ride the alert metadata.
Enroll the people you know, and a detection becomes a named alert. On the edge node, faces are aligned, embedded, and matched against your private gallery — so “person detected” turns into “a known face at the front door.” Enrollment and matching run entirely on-prem: no cloud service, no faces ever leaving your network.
Rolling segment buffer with event-triggered MP4 clip extraction, plus a low-resource JPEG ring-buffer mode. Optional one-pass GPU H.265/H.264 re-encode shrinks footage with on-the-fly playback transcode for any browser.
Direct H.264 RTP-to-WebRTC live view with MJPEG fallback and live sound on any camera codec (G.711 pass-through, AAC→Opus transcoding). Two-way audio talk-back and press-and-hold PTZ let operators respond, not just watch — across paged grids from 1×1 to 4×4.
The same idea as the NVR, but for everything that is not a camera: door contacts, motion, temperature, smoke, leaks, power meters and Modbus/SunSpec inverters, over an embedded MQTT broker — plus an opt-in, read-only LAN scan that finds what is already on the network. A deadbanded store keeps history small, threshold and correlation rules raise alerts, and scenes, schedules and a visual flow canvas automate the response. Every command back to a device is forced through one bounded, rate-limited, fully audited chokepoint.
Badge readers speak OSDP over RS-485 or IP, and every badge-in is decided on the appliance itself against people, groups, schedules and holidays — so a cut uplink never means a door that will not open. Site-wide lockdown, a duress PIN that opens the door while silently raising the alarm, and an append-only activity log come as standard, and the strike only ever fires through one audited chokepoint.
The rule no single appliance can write on its own: motion on a camera AND a door opening AND no badge accepted, inside the same window, at the same site. The control plane arms on the first signal, waits out a grace period for the innocent explanation, and only then raises one correlated incident instead of three unrelated alerts.
Recordings, snapshots, and training images are AES-256-GCM encrypted on disk; factory reset crypto-erases by destroying the key and multi-pass shreds deleted footage. The key is wrapped by an OS keystore (DPAPI / systemd-creds) or a portable passphrase with an exportable recovery escrow — and a passphrase-encrypted .mmbackup moves cameras, rules, and settings between hosts.
No ML jargon, no dataset wrangling. A guided wizard teaches the camera a new skill — name it, pick a goal (recognize a new object, tell good from bad, or spot anything unusual), show a few examples, check the accuracy, and turn it on. It builds and trains a real custom model behind the scenes, then hot-swaps the live detector.
One feed across AI detection, camera and machine health, and login security — with per-event acknowledge, annotated screenshots, and in-page clip playback. Filter the feed down to a single source camera to isolate one view’s events. Route to webhook, Telegram, or MQTT.
Authenticated UDP multicast discovery + single-parent adoption with a short-lived claim code. Nodes enroll for a fleet-CA certificate and serve a mutual-TLS management channel — no inbound ports, no cloud broker. Each node’s certificate auto-renews on an operator-gated schedule, so fleet trust stays current and no forgotten node lapses.
The control plane puts your whole estate on a map — sites and buildings pinned by location, with every edge node and its cameras layered on top. Drop into a building to design its floor plan in-app, place each camera on the plan with a field-of-view coverage cone, and see one building’s cameras aggregated across several nodes. Click any camera to locate it or open its live view straight from the map.
A daily — and optionally weekly — digest reads the whole estate and reports what changed: event-volume swings, a node breaking its own learned baseline, outages, expiring certificates, noisy sources, sensitive audit activity, even a suggested rule for after-hours activity nothing covers yet. Every number is computed in plain Go, so a digest finding is never a hallucination. An optional language model — your own endpoint, or a supervised sidecar you can carry in on a USB stick — adds an "ask the fleet" chat that cites the records it used and answers "how do I…" straight from the built-in manuals. It runs on your hardware or not at all.
The unified event feed becomes insight — KPI tiles, detections over time, and per-category, per-camera breakdowns with a range selector and auto-refresh. It goes past charts: an activity heatmap by hour and weekday, an expected-range band that flags when a camera breaks its own normal pattern, spike and "unusual silence" anomaly alerts, a per-camera reliability scorecard, and a notification noise ratio. Aggregation runs server-side, so it works on SQLite or a full database engine.
Search your footage by what the cameras actually saw. Every detection is coalesced onto a searchable timeline — filter by camera, date range, one or more object types at once, and confidence, then jump straight to the exact moment in the recording, with the object boxed on a preview thumbnail. Export the results to CSV or PDF. It taps the live detector's own output, so there is no second video pass.
One sign-on across every app, from a single pure-Go binary that never leaves your intranet. Local accounts alongside LDAP / Active Directory, Kerberos SPNEGO desktop SSO and generic OIDC providers, with group-to-role mapping and per-app, per-page access control. TOTP and WebAuthn security keys, password policy, lockout and step-up re-authentication guard the front door; an immutable audit trail, live session administration and passphrase-encrypted backup/restore keep it accountable.
A first-run wizard walks setup end to end in every app, a capacity estimator sizes your host, and machine-health monitoring self-heals — overwriting the oldest footage before the disk fills. Each app also carries its own printable four-language user manual compiled into the binary, readable offline from the sign-in screen. ffmpeg, the Python AI runtime, and app updates all install from inside the app.
How it works
Authenticated ONVIF discovery and manual probe find cameras on your LAN, an enrollment window or read-only network scan finds sensors, and OSDP readers announce themselves on the bus. Saved devices persist in a local SQLite database.
On-device YOLO inference runs your detection rules — presence, crowd, intrusion, line crossing, or plate recognition — against the live frames, while sensor thresholds and badge decisions are evaluated on the appliance itself.
Continuous NVR recording rolls in the background and extracts an MP4 clip the moment a rule fires, encrypted on disk; telemetry and every access decision land in their own append-only history.
Alerts land in the unified feed with an annotated snapshot and clip, correlate across cameras, sensors and doors at the control plane, and fan out to webhook, Telegram, or MQTT destinations you choose.
Hardware
The camera node is the hungry one: it runs from a pocket-sized appliance to a GPU server, and you choose the tier. The sensor hub, door controller, identity broker and control plane are single pure-Go binaries that fit comfortably alongside it. (Camera counts are rough; the built-in capacity estimator measures the real number for your host.)
Raspberry Pi 4/5 · 4 cores · 2–4 GB
~1–4 cameras
Live view + recording for a couple of cameras; AI is native motion or YOLO-nano at a relaxed interval. SQLite, CPU software decode, no GPU.
Mini-PC / NUC · 6–8 cores · 8–16 GB
~6–16 cameras
Continuous recording with real-time YOLO detection at the 2 s default. An entry NVIDIA GPU lifts camera count and keeps AI real-time.
Server · 12+ cores · 32 GB+ · NVIDIA GPU
20+ cameras
GPU-accelerated detection, in-app model training, HEVC NVENC compression, and long retention. Server DB engine + Redis at scale.
See it in action
These screens are MyMataSan, the camera node: live multi-camera grids, detection rules drawn on the real frame, guided camera teaching with no ML jargon, continuous NVR recordings, event analytics, encrypted backups, and version & health. Every app in the suite shares the same shell — in the browser, in four languages.
Use cases
Train a custom model to flag product defects, missing parts, or PPE gaps on the line — private, on-prem, with no per-seat AI fees.
Crowd and intrusion alerts, plus license-plate recognition for forecourt and drive-through monitoring.
Watch loading docks and yards — vehicle and person detection with line-crossing at gates, all on edge hardware.
After-hours intrusion detection with secure, encrypted recording that stays on-site.
Badge readers, access schedules and site lockdown decided on-site, with the camera that watched the door and the sensor that felt it open reporting into the same timeline.
Spot animals or intruders across land with poor or no connectivity; detection and recording run entirely locally.
After-hours movement and fall-style alerts with footage that never leaves the building — privacy by default.
Zone and line-crossing rules for restricted areas and equipment, on rugged hardware where uplinks are unreliable.
A control plane adopts many edge nodes over the LAN — cameras, sensors and door controllers alike — maps every site and building, and relays live view back to operators, with per-building floor plans that place each camera and its coverage.
The platform
r450k is a modular platform. mymatasan is the edge camera node, myiotsan the edge sensor hub, and mypintusan the door controller; the control plane adopts all three and correlates events across them, and shared identity ties it together.
Flagship · in active development
The standalone edge camera & video-intelligence node: ONVIF, AI detection, NVR, WebRTC live view, encryption, and LAN pairing.
Sensor hub · available now
The NVR, but for sensors: door contacts, motion, temperature, smoke, leaks, power meters and Modbus/SunSpec inverters over an embedded MQTT broker — with rules, alerts, scenes and schedules, a visual flow canvas, safe actuation, and telemetry history.
Access control · in development
The door controller: OSDP badge readers over RS-485 or IP, with people, groups, schedules, holidays, site lockdown and a duress PIN all decided on the appliance itself, so a cut uplink never means a door that will not open. Adoptable into the fleet like any other node — not yet available to download.
Control plane · available now
The fleet control plane that discovers, adopts, and manages camera, sensor and door nodes alike — mapping every site, building, and node with per-building floor plans and camera coverage — and correlates events across them: motion on a camera AND a door opening AND no badge accepted. It also writes the daily fleet digest, answers questions about the fleet, and renders printable PDF reports.
Identity · available now
The single sign-on front door: one federated identity across every app, with local accounts, enterprise LDAP / Active Directory, Kerberos SPNEGO desktop SSO, and generic OIDC providers — plus TOTP and WebAuthn multi-factor, group-to-role mapping, per-app access control, and an immutable audit trail. A single pure-Go binary that runs fully on your intranet, no egress.
Pricing
Personal and non-commercial use is free — forever. Businesses keep the project alive with a simple commercial license. This is early pricing; we’ll tune it together as things grow.
Prices in USD. Non-commercial use stays free. Not sure which fits? Reach out and we’ll figure it out together.
Get the apps
Run one node at a single site, or a whole fleet from one control plane. Each is a single install with the web UI and a default config bundled. (mypintusan, the door controller, is still in development and not downloadable yet.)
Free for personal and non-commercial use — individuals, non-profits, education, and research. Commercial or in-business use, and any reselling, needs a commercial license.
Loading the latest release…
r450k brings cloud-grade camera AI, sensor automation and access control to hardware you already own, with privacy as the default rather than an upgrade.
Explore the features