Chief Developer & Designer: Rob Branting

AI‑Stalker is the next evolution of AutoPTZ — now reborn as a multi‑node, AI‑amplified, fail‑safe security platform.
High‑energy, high‑reliability, zero‑excuses.
We hack the boredom out of security — legally, ethically, and with consent.

AI‑Stalker is for authorized security and automation use only.
No unauthorized surveillance. No device hijacking. No mischief.
If your use case needs a lawyer, it’s not welcome here.

1. What AI‑Stalker Is
2. Current Capabilities (Today)
3. Planned Capabilities (Roadmap)
4. Feature Matrix
5. Architecture at a Glance
6. Advanced Behaviors & Protocols
7. Failover & “Next‑in‑Line” Failsafe
8. Hive Mode & Idle Resource Pooling
9. AI Engines & LLM Strategy
10. Device & Smart‑Home Integration
11. Observability & Operations
12. Deployment Topologies
13. Installer & Service Mode
14. Installation (Developer)
15. Implementation Blueprint
16. License

AI‑Stalker is a distributed, AI‑enhanced local-first NVR + security orchestrator that scales from a single PC into a hardened multi‑node “security hive”—built to keep your video and context in-house, keep your life calm, and keep contributors honest.

Blacklisted Binary Labs edition: local-first power, consent-first discipline, and zero paywall nonsense.

• 100% free in all ways: no paywalls, no forced upgrades, no “pro-only” core capability locks.
• Local-first AI: sensitive feeds and metadata stay on your network by default (cloud only if you explicitly choose it).
• Consent-based engineering: for authorized security + automation use only.
• Auditable design: if it matters, it’s inspectable—no “trust me” magic.

AI‑Stalker is designed to:

• Run fully local (offline‑first), with optional hybrid cloud AI.

• Coordinate multiple computers with automatic failover (leader/failsafe chain).

• Pool idle compute from secondary nodes to boost inference only when idle, with safe auto-retreat.

• Deliver advanced monitoring: PTZ automation (VISCA), AI tracking, identity workflows (trusted/ignored + high-risk priority), and event summarization (who arrived/left, who they interacted with, what they brought/left with, and behavior/context tagging where feasible).

• Expand with authorized repurposed user-owned devices (cameras/mics today; repurposed smartphones later).

• Provide advanced resilience: cluster health + automatic continuity if a node goes down.

• Merge cameras, microphones, and sensors into a single brain.

These features exist in the current codebase:

• ✅ Live camera feeds (USB and NDI; RTSP under development)
• ✅ Facial recognition & tracking
• ✅ Automated PTZ movement via VISCA (network + USB)
• ✅ Cross‑platform runtime (Windows/macOS)
• ✅ Qt UI for control and monitoring

These features are planned and detailed in the blueprint:

• 🧠 Multi‑LLM orchestration (local + hosted AI providers)
• ⚡ OpenVINO / Triton inference for CPU‑optimized AI
• 🧩 Pluggable AI backends (InsightFace, OpenFace, Kornia)
• 🛰️ Failover node chain with automatic takeover (1–60 min delay)
• 🐝 Hive mode: use idle computers for AI compute + storage
• 🧰 Smart‑home device integration (lights, locks, sensors)
• 🧾 Observability pipeline for logs + events across all nodes

• Local‑first AI: keep sensitive video on your network.
• Failover baked in: not a bolt‑on, not a manual script.
• Hive compute: idle machines become a privacy‑friendly AI cluster.
• Protocol agnostic: USB/NDI/RTSP/ONVIF with no vendor lock‑in.
• Operator‑first UI: built for quick response, not endless settings.
• Ethical guardrails: consent‑based device onboarding by design.

[Cameras/Mics/Sensors]
        │
        ▼
[Ingest + Normalization] → [AI Pipelines] → [Events & Alerts]
        │                      │                │
        ▼                      ▼                ▼
     [Storage]           [LLM Orchestrator]  [Automation]

[Control Plane: Raft + Memberlist]
[Data Plane: Zenoh + NATS]
[Sync: Syncthing / DRBD]

flowchart LR
  A[Cameras/Audio/Sensors] --> B[Ingest]
  B --> C[AI Pipelines]
  C --> D[Alerts & Actions]
  C --> E[Event Bus]
  E --> F[Automation/Smart Devices]
  B --> G[Storage]

AI‑Stalker is designed around intentional, auditable behaviors and standard protocols:

• NDI (live video ingest)
• USB Video (local devices)
• RTSP / ONVIF (planned for IP cameras)
• VISCA (PTZ control)
• NATS / Zenoh (event and telemetry)

• Alert Cascade: one sensor triggers higher sensitivity on nearby nodes.
• Zone‑aware escalation: raise alert thresholds based on time and location.
• Multi‑camera correlation: fuse detections across nodes.
• Audio‑driven wake‑up: trigger camera recording on sound patterns.

• Consent‑based device enrollment with explicit approval.
• Encryption in transit and at rest for logs, embeddings, and metadata.
• Audit trails for configuration and admin actions.
• Minimal data retention policies by default.

AI‑Stalker is built like a relay race for reliability:

• Leader election via Raft/Dragonboat
• Membership tracking via Memberlist gossip
• Failover window configurable from 1–60 minutes
• Config + model data sync (Syncthing / optional DRBD)

1. Primary node goes offline (unexpected shutdown).
2. Cluster detects missing heartbeat.
3. Next node enters standby‑promotion state.
4. VIP/DNS takeover, services resume.
5. No manual reconfiguration required.

AI‑Stalker can borrow idle machines like a polite vampire:

• ✅ Idle detection via CPU + input activity thresholds
• ✅ Auto‑retreat when users return
• ✅ Compute pooling for AI inference
• ✅ Storage pooling for redundancy

Local‑first, cloud‑optional:

• OpenVINO for CPU‑optimized inference
• Triton Inference Server for multi‑model hosting
• InsightFace/OpenFace for face embeddings
• Kornia (Rust) for accelerated CV filters

LLM routing policies:

• Local model by default
• Cloud only if explicitly allowed
• Policy‑driven token budgeting + privacy constraints

AI‑Stalker turns your environment into a coordinated defense system:

• Home Assistant bridge for smart devices
• Zigbee support via zigpy (optional)
• Audio alerts with Rhasspy
• Device onboarding only with explicit user authorization

Because invisible problems are the worst kind:

• Netdata for live node performance
• Vector for log pipelines
• NATS for alert delivery guarantees

1) Solo Node: one PC, all features local.
2) Primary + Failsafe: a standby node that takes over on outage.
3) Hive Cluster: multiple nodes, pooled compute, shared storage.

• Role profiles: primary, failsafe, hive worker.
• Device profiles: camera/mic/sensor templates.
• Policy profiles: privacy, retention, AI routing.
• Idle policies: thresholds for compute lending.

A Windows installer blueprint is provided at:
installer/windows/ai-stalker.iss

Planned installer features:

• 32/64‑bit unified setup
• Service installation (runs without user login)
• Failsafe node configuration during setup
• Desktop icon + auto‑start options

AI‑Stalker builds on the current AutoPTZ codebase.

• Python 3.7+
• Windows or macOS (Linux is untested but likely viable)

pip install cmake
pip install dlib
pip install -r requirements.txt
python startup.py

For the full technical roadmap, see:
AI_STALKER_IMPLEMENTATION_BLUEPRINT.md

Q: Is this a stalker tool?
A: No. The name is branding. The product is for lawful, consent‑based security.

Q: Can it run without the cloud?
A: That’s the goal. Local inference is the priority.

Q: Will it run as a Windows service?
A: The installer blueprint includes service mode.

See LICENSE.md.

We don’t do evil. We do engineering.
We don’t stalk people. We protect spaces.
We don’t cheat the system. We out‑design it.
We wear black because it hides the coffee stains — not because we’re the villains.