One Camera Sees It. Four Systems Confirm It. Edge AI + Observability Fusion at Mumbai & Chennai hub sites AI detection ISE identity BMS occupancy Historical patterns FUSION GATE multi-source validation Automated action <500 ms end to end <5% false positives
Edge AI · Jul 2026

Edge AI + Observability Fusion: Trusting the Machine to Act

Rajmohan M
Principal Consultant, UC & Contact Center
4 min read · July 2026
AI-Assisted Documentation

Edge AI that only watches is easy. Edge AI you trust to act — to lock a door, page a guard, adjust a building — is a different engineering problem, because a single camera's opinion isn't enough to act on. This design solves that with fusion.

The deployment runs distributed AI inference at the Mumbai and Chennai hub sites, built as an illustrative enterprise scenario. Every AI detection is validated against three independent sources before anything automated happens — and that fusion is what brings false positives from the 15–30% typical of camera-only edge AI down to under five percent.

01. The deployment at a glance

Four edge nodes, two sites, one platform. Mumbai deploys first as the pilot, Chennai follows using the identical build.

MUMBAI · weeks 1–8 (pilot) 2 edge AI nodes · primary + standby 135 cameras · all three use cases UCS XE9305 + XE130c M8 + NVIDIA L4 24GB CHENNAI · weeks 9–12 (rollout) 2 edge AI nodes · identical build 135 cameras · all three use cases weeks 13–16: production hardening UC1 · Physical security perimeter & intrusion detection UC2 · Building automation occupancy-driven HVAC via BMS UC3 · Safety compliance zone & PPE monitoring

02. One new layer, everything else reused

The cameras, the Catalyst 9300/9500 network, and the Honeywell BMS were already there. The design adds a single new layer — edge AI compute in the same IDF room — which is why the camera-to-inference path is 4 ms instead of the 15–20 ms of a separate datacenter.

Cameras & sensors existing · SGT-70 Catalyst 9300 / 9500 existing network · reused Edge AI layer XE130c M8 + L4 GPU · new same IDF room · SGT-95 Honeywell BMS occupancy-driven HVAC Security team validated alerts only 10G Camera to inference: 4 ms in the same IDF room — versus 15–20 ms if the GPU lived in a separate datacenter.
Existing · reused New edge AI layer

03. Detection to action in under 500 ms

The whole point of edge placement is the timeline. From the frame leaving the camera to an automated action firing, everything — inference, validation across four sources, and the action itself — completes in under half a second.

Frame capture camera → IDF · 4 ms GPU inference NVIDIA L4 · on-site Multi-source validation ISE · Splunk · TE · AppD Automated action alert · lock · adjust t = 0 t < 500 ms Millisecond-budgeted end to end — the design documents the timing at every hop.

04. Why fusion beats a smarter camera

Camera-only edge AI acts on one opinion, and pays for it with 15–30% false positives. Here, the AI detection is one vote of four: identity from ISE, building context from the BMS, and historical patterns from the observability stack all confirm or veto before any action fires.

AI detection “person at perimeter door” ISE · identity check badge seen on this door in last 60 s? BMS · building context zone occupied? after hours? History · pattern check is this normal for this hour & door? 3 of 3 agree? Act automatically <5% false positives yes Route to a human alert with full context no The machine only acts alone when independent systems agree — that is the difference between 5% and 30% false positives.

05. Privacy is an architecture decision

Two rules are wired into the design rather than written into a policy: no video ever leaves the site, and no facial recognition runs anywhere. Video is inferred on-site and discarded; only event metadata travels — over MACsec, between segmented zones.

SITE BOUNDARY — video never crosses this line Camera video SGT-70 segment Edge AI node infer & discard · SGT-95 no facial recognition MACsec metadata only Observability stack Splunk · ThousandEyes · AppD FTD · XDR · ServiceNow events, never frames A privacy promise you can audit in the topology is stronger than one written in a policy document.

Where AI fits in

The design spans four chapters and nine platform integrations, and Claude helped me draft it at scale — the per-use-case detection specifications, the millisecond timing tables, and the week-by-week roadmap — while the architecture decisions, the fusion logic, and the privacy boundaries are mine. Every page passed the same strict build gate as the rest of this portfolio.

Explore the full design

Everything above unfolds into complete documentation — the business case, the fusion architecture with full detection timelines, all three use case specifications, the zero-trust segmentation design, and the sixteen-week implementation roadmap. Built as an illustrative enterprise scenario so the structure and decision logic are fully visible.

Full Documentation Site

Edge AI + Observability Fusion Architecture

The complete Mumbai & Chennai edge AI design, published as a browsable site.

Open the Edge AI Guide →
AI-assisted disclosure: This article and the documentation site it references were produced with AI assistance (Claude, Anthropic) under the author's technical direction, as part of the AbhavTech knowledge-sharing portfolio. Content is illustrative and intended for learning; performance figures are design targets of the documented scenario — validate all designs independently before applying them to production environments.