Private 5G Edge-AI Capacity Console

AGV fleet growth, worst-cell latency, and edge GPU bottlenecks under a 20 ms control-loop budget. Static evidence only: seeded simulation, derived capacity metrics, sensitivity, benchmarks, and explicit boundaries.

Portal (cards view)One-page brief README Benchmarks

Decision

100 AGVs

Safe simulated capacity at the 20 ms p95 budget.

Failure point

120 AGVs

First tested expansion that crosses the budget.

Bottleneck

Assembly edge GPU

Saturation appears first in the seeded outputs.

Confidence

Seeded + deterministic

Replay, benchmark, and metrics artifacts regenerate.

Boundary

Simulation only

Seeded evidence; no live private 5G network.

Operational Decision Summary

How far can the AGV fleet grow before the control loop breaks?

Recommended ceiling

100 AGVs

First unsafe test point

120 AGVs

Control-loop budget

20 ms

Recommended simulated fleet ceiling	100 AGVs
First tested unsafe expansion point	120 AGVs
Binding bottleneck	Assembly-zone edge GPU saturation appears first in this seeded simulation.
Primary metric	worst-cell hourly p95 latency
Control-loop budget	20 ms p95
Decision basis	seeded simulation + generated plots + budget sensitivity + multi-seed stability
Operator recommendation	Cap expansion at 100 AGVs until live telemetry validates headroom.

Problem -> What I Built -> What I Found -> What I Would Do

The operator story in one pass

Problem

A factory can add AGVs and still see healthy-looking private 5G metrics while edge inference latency quietly becomes the real constraint.

What I Built

I built a reproducible telemetry-intelligence pipeline that joins private-5G-style KPIs, AGV fleet load, edge GPU pressure, schema validation, quarantine handling, deterministic simulation, benchmarks, and a static decision dashboard.

What I Found

The simulated floor stayed under the 20 ms p95 budget through 100 AGVs. At 120 AGVs, the budget broke. Assembly-zone edge GPU saturation appeared first.

What I Would Do

I would cap expansion at 100 AGVs, rebalance assembly-zone workload, reserve edge GPU capacity before testing 120 AGVs, and validate with live RAN, GPU, and factory telemetry before any real deployment decision.

Bottleneck Attribution

What appears to break first, and why?

Candidate	Finding	Evidence	Decision implication
Edge GPU pool	Assembly-zone edge GPU saturation appears first in the seeded simulation.	Assembly zone hits the saturation knee first and carries 42% of the fleet.	Reserve or add edge GPU capacity before testing 120 AGVs.
Radio / private 5G layer	Not shown as the first bottleneck in this seeded simulation.	The latency break is tied to edge GPU pressure and worst-cell p95 behavior, not a proven radio-only failure.	Do not assume radio-capacity spend is the first fix.
Zone-level fleet distribution	Assembly carries the heaviest load.	42% assembly, 34% paint, 24% warehouse.	Rebalance routing or workload before expanding fleet size.
Worst-cell latency budget	100 AGVs remains under the 20 ms p95 budget, while 120 AGVs crosses it.	18.39 ms at 100 AGVs and 23.40 ms at 120 AGVs.	Keep the simulated ceiling at 100 AGVs until validated.
Data pipeline health	Evidence path is reproducible and tested.	Schema validation, quarantine path, benchmarks, deterministic regeneration, tests/CI.	The analysis is reproducible, but it is still not live production telemetry.

Operator Action Plan

What the operator should do next

Action	Priority	Why	Evidence	Boundary
Cap expansion at 100 AGVs	P0	Last tested fleet size under the 20 ms p95 budget.	100 AGVs = 18.39 ms p95.	Seeded simulation, not production approval.
Validate with live RAN + edge GPU telemetry	P0	Current evidence is generated from seeded scenarios.	Boundary section and readiness checklist.	Required before any real deployment decision.
Rebalance assembly-zone AGV routing/load	P1	Assembly zone carries the heaviest load and saturates first.	42% assembly fleet share.	Validate against real routing and factory layout.
Reserve or add edge GPU capacity before testing 120 AGVs	P1	Edge inference pool appears to be the limiting layer.	120 AGVs crosses 23.40 ms p95.	Confirm with live GPU utilization and queueing telemetry.
Re-run scenario at 15 ms, 20 ms, and 25 ms budgets	P1	Control-loop deadline changes the allowable fleet size.	15 ms = 80 AGVs, 20 ms = 100 AGVs, 25 ms = 120 AGVs	Budget choice must come from actual control-loop requirements.

Evidence vs Boundary

What is demonstrated, and what is deliberately not claimed

Evidence demonstrated

Seeded AGV fleet simulation
Worst-cell p95 latency budget check
Edge GPU load / zone capacity signal
Zone-level load signal
Schema validation and quarantine behavior
Deterministic artifact regeneration
Benchmark and CI validation

Boundary preserved

No live private 5G network
No vendor RAN integration
No MES/SCADA/PLC connector
No real AGV fleet telemetry
No production deployment claim
No safety certification claim

Operator Console Readiness

Implemented evidence versus deployment boundaries

Seeded simulation	PASS
Deterministic regeneration	PASS
Schema validation	PASS
Quarantine path tested	PASS
Benchmark artifacts	PASS
Live private 5G telemetry	NOT IMPLEMENTED
Vendor RAN integration	NOT CLAIMED
MES/SCADA/PLC integration	NOT IMPLEMENTED
Safety certification	NOT CLAIMED
Production deployment	NOT CLAIMED

Manufacturing

How many AGVs can the floor add before the 20 ms control-loop budget breaks?

A factory floor running edge-AI vision on AGVs over private 5G. Ops wants to grow the fleet. The radio isn't the bottleneck — the edge GPU pool feeding inference to the AGV cameras is. I swept fleet size from 20 to 160 AGVs across three URLLC cells (assembly, paint, warehouse) and watched worst-cell hourly p95 latency.

Max fleet under budget

100 AGVs

Budget

20 ms

First over-budget fleet

120 AGVs

Worst-cell p95 latency vs fleet size with 20 ms budget line — Worst-cell hourly p95 latency stays under the 20 ms budget through 100 AGVs. At 120 AGVs the assembly-zone cell has saturated; latency reaches 23.4 ms.

Edge GPU load by zone — Assembly zone hits the saturation knee first — it carries 42 % of the fleet versus 34 % paint and 24 % warehouse.

Per-sample latency over the fleet sweep — Every per-sample latency reading. Blue stays under budget; red goes over. The cloud climbs through the back half of the sweep.

Headline reference — Headline view (same plot, larger above).

Per-fleet detail

Fleet	Worst-cell p95 (ms)	Mean p95 (ms)	Mean edge load	Budget
20	7.62	7.08	0.27	ok
40	7.53	7.28	0.34	ok
60	9.29	8.10	0.40	ok
80	14.56	10.67	0.46	ok
100	18.39	14.17	0.52	ok
120	23.40	18.47	0.60	over
140	27.78	22.18	0.67	over
160	31.51	24.82	0.73	over

If you change the budget

Latency budget	Max fleet
15 ms	80 AGVs
20 ms	100 AGVs
25 ms	120 AGVs

Full business case (sensitivity + four-seed stability) →

Secondary portability scenario

How early does the line catch a bioreactor contamination event?

Secondary portability scenario, not the headline use case. This is seeded diagnostic evidence that the pipeline shape can support another vertical. It is not real pharmaceutical plant telemetry.

To show the pipeline isn't single-use, I pointed it at a non-manufacturing vertical: a pharma cleanroom bioreactor with slow sensor drift through the shift and one ~75-minute contamination event injected late in the day. The detector is an online z-score on the edge-load × latency signal, using a trailing 1-hour baseline with a 10-minute gap before the test sample.

Lead time

5 min

Precision

1.00

Threshold

z = 3.0

Detector z-score across the day with event window shaded — The detector's z-score sits in the noise band through the drift hours, then spikes at the event onset above the threshold line.

Edge × latency signal across 24 hours — The underlying signal — edge inference load × latency — climbs sharply at the event window (shaded red).

Detector z-score reference — Detector z-score (same plot, larger above).

Threshold sweep

z-threshold	Detected	Lead time	TP windows	FP windows	Precision
2.0	yes	5 min	4	3	0.57
2.5	yes	5 min	3	2	0.60
3.0	yes	5 min	3	0	1.00
3.5	yes	5 min	2	0	1.00
4.0	yes	5 min	2	0	1.00

Same result across four seeds

Seed	Detected	Lead time	Precision
42	yes	5 min	1.00
7	yes	5 min	1.00
17	yes	5 min	1.00
99	yes	5 min	1.00

Full business case (threshold + seed sensitivity) →

Measured pipeline benchmarks

Not FPS. What's actually meaningful for a telemetry layer.

This isn't an inference workload, so FPS and GPU utilisation would be staged. The benchmarks that are meaningful for a telemetry-intelligence layer: ingest throughput, schema-quarantine throughput under corrupted input, end-to-end regeneration wall-clock, and determinism. All measured.

Ingest

10,455

rows / sec

Quarantine

168,237

rows / sec @ 5% corruption

Regeneration

16.1s

both scenarios + cases + portal

Determinism

byte-identical

SHA-256, two runs at seed 42

Full methodology + stage breakdown →

What I tested, what I didn't

Every number on this page comes from a seeded simulation. I didn't connect to a live private 5G network, vendor RAN, vendor edge-AI platform, MES, SCADA, PLC, or a real AGV fleet. What's portable is the shape — the schema, the scenario pattern, the way each answer comes with its sensitivity story attached.