← all challenges

// challenge brief · hack a ton 2026

Watchful

Tell a camera what to watch for in plain English; when it happens, the agent acts.

// Proposed by
ThePlace
// Industry
Hospitality / smart venue
// Difficulty
🟡 Intermediate

★ ThePlace is open to running the winning build on-site.

// the problem

Security and venue cameras see everything and understand nothing. Acting on what they see means a human watching a wall of screens. What if you could just tell the camera what matters, in your own words, and let it take the action for you?

// your mission

You're given an IP address and login for an existing Hikvision IPCAM/NVR. Build an AI agent that monitors the live feed and takes actions based on natural-language conditions the user defines.

// how it works

  1. 1Perceive: the agent watches the live image stream.
  2. 2Understand: the user writes conditions in plain language.
  3. 3Act: when a condition is met, the agent triggers the matching action.

Example prompts:

  • "Every time someone sits on the underwater chair in the jacuzzi and raises their hand, turn on relay 1 for 5 minutes."
  • "If there's nobody in the image for more than 15 minutes, turn relays 1, 2, 3 OFF."
  • "Notify me whenever more than 10 people are in the event hall."

Available actions: trigger relays (on/off/duration), send notifications (webhook with alert details), log events.

// what we provide

  • IP address + credentials for a live Hikvision IPCAM/NVR.
  • The available relay/action interface.

// suggested approach

  • Inference: for model access, we suggest [LLMok](https://llmok.app). Use code AMBASADA26 for 50% off.
  • Pull frames over RTSP; control relays/I-O via Hikvision ISAPI (HTTP).
  • A vision-language model turns each natural-language condition into a checkable predicate over frames; an agent loop evaluates and fires actions.
  • Floor version: one condition, one relay, polling a few frames per second. The hard part isn't detecting. It's not firing on shadows.

// how you'll be judged

  • Does it correctly detect the conditions?
  • Does it trigger the right actions?
  • How fast does it respond?
  • How creative and reliable is the implementation (low false-trigger rate)?

// stretch goals

  • Multiple simultaneous conditions across multiple zones.
  • Confidence thresholds and debouncing to suppress false positives.
  • A simple UI to add/edit conditions in natural language.

// deliverables

  • Git repo: code + README + run instructions.
  • Demo: live on the provided camera, or on recorded footage.
  • 5-minute pitch: architecture, decisions, trade-offs.