Centinela Agent

In Peru's electronic security sector, companies that install CCTV, access control, and fire detection systems at retail sites are only paid after submitting a formal compliance report with visual evidence for every device. Today, an assistant spends 12+ hours per site manually collecting photos from chaotic group chats, pairing before/after evidence, and writing observations — a critical bottleneck that delays cash flow. Centinela Agent integrates directly into the messaging apps field technicians already use on-site. As crews document their work by sending photos and captions, the agent ingests each image through a CPU preprocessing pipeline — compression, perceptual hashing, blur detection — clusters near-duplicates, and queues only the sharpest representative per group for GPU inference. A Gemma 4 31B multimodal model served via vLLM on an AMD Instinct MI300X (192 GB HBM3) classifies each photo: equipment category, intervention state (in-progress vs. completed), floor plan location, OCR data, and technician observations. Anomalies trigger a Human-in-the-Loop confirmation flow with the field supervisor. When a project closes, a structured compliance report is generated automatically. What took 12 hours now takes under 15 minutes of human review.