Problem: With 50+ networked machines running simultaneously, early testing revealed cascading failures — one room crashing could ripple into adjacent rooms sharing network resources. Sensor drift on the LiDAR arrays caused false triggers in high-traffic areas.

Fix: Isolated each room onto its own VLAN with independent watchdog processes and auto-restart sequences. Implemented health-check heartbeats so the central dashboard could flag and reboot individual machines without affecting the rest of the venue. For LiDAR drift, built calibration routines that run during overnight downtime and added dead-zone filtering to ignore edge-case noise. The system now self-heals without staff intervention.

Problem: The GitHub API rate limits nearly throttled The Core during peak hours. With hundreds of attendees searching their handles simultaneously, the system started queuing requests and the particle visualization stalled.

Fix: Implemented an aggressive local cache layer — once a handle was looked up, the data was stored locally for the rest of the conference. Added a pre-fetch queue that pulled popular/trending handles during off-peak. The result was sub-second lookups for returning queries and graceful degradation for new ones during spikes.

Problem: The elevator API was undocumented and intermittent — it would drop connection unpredictably, causing the generative visuals to freeze mid-animation. The building's IT team had no technical documentation on the elevator data protocol.

Fix: Reverse-engineered the elevator data feed by logging raw packets over several weeks. Built a resilient connection handler with automatic reconnection, data smoothing (to fill gaps during dropouts), and a graceful fallback to ambient generative content when the elevator feed goes offline. The wall never shows a frozen frame — it degrades into a beautiful ambient state instead.

Problem: During early load-in testing, the wireless headphone system dropped audio when too many guests clustered in the transition zone between rooms. The room-detection beacons had overlapping ranges, causing the app to rapidly switch audio channels — creating a jarring "flip-flop" effect.

Fix: Added a debounce layer to the room-detection logic — the system required a sustained signal from a new room's beacon before switching audio channels. Also adjusted beacon power levels and placement to create cleaner handoff zones. The transition became seamless, and guests never noticed the switch.

Problem: During rehearsals, the game data API would occasionally send burst packets — hundreds of events in a single frame — when a major in-game moment happened (like a mass elimination). This caused the generative engine to spike and drop frames on the broadcast feed.

Fix: Built a throttle/smoothing layer between the data ingestion and the generative engine. Major events were queued and spread across a short animation window (250ms) instead of all hitting in a single frame. This preserved the dramatic visual impact while keeping the render pipeline stable. The broadcast feed ran clean for the entire 3-day event.

Problem: Kitchen timing is inherently unpredictable — courses could arrive 5 minutes early or 15 minutes late depending on the evening. The original show control was timecode-based, which meant the visual scene would either finish too early (leaving dead air) or get cut short when the next course arrived.

Fix: Redesigned the show control to be cue-based instead of timecode-based. Each course had a start cue triggered by the kitchen team (via a simple tablet interface), and the content was structured with extendable ambient loops that could stretch or compress gracefully. Transitions always looked intentional, regardless of when the kitchen called the next course.