Skip to content

Customers — Production Deployments

Four named enterprise deployments. Measured outcomes. No logo wall.

FGCV ships in production at global manufacturers and logistics operators where a missed defect or a silent model drift is unacceptable. Below: four deployments, each annotated with the stack used, the runtime region, and a single quantified result — then the cross-customer benchmarks that aggregate 312 deployments in 2024.

  • Scope4 verified deployments
  • IndustriesIndustrial · Automotive · Logistics · Retail
  • RegionsEU · NA · APAC
  • UpdatedQ1 2025

Four production deployments, indexed.

Each row is a single environment: the modules shipped, the runtime region, the headline metric. No marketing copy — the stack sidebar speaks for itself.

  1. Defect detection on PCBA lines

    From stitched-together scripts to a single versioned pipeline.

    Siemens’ CV team replaced an in-house stack of training scripts, a third-party label tool, and a custom inference service with FGCV’s label-to-edge pipeline. Defect taxonomy now lives in a versioned ontology, and the active learning loop surfaces only the ambiguous frames to human reviewers — cutting annotation spend across the PCBA lines.

    Result 71% reduction in labeling cost on the manufacturing defect dataset.
  2. Weld inspection across 3 plants

    One model, three plants — shipped in a quarter, not six months.

    BMW’s quality group needed a single weld-inspection model deployed across three European plants with differing camera rigs and lighting. FGCV’s evaluation harness produced per-site regression scores before every rollout, and the edge runtime kept inference predictable on shared T4 hardware. Integration timelines compressed because the deployment artifact is the same one the team already reviews in staging.

    Result Time-to-production cut by 4.8x versus the prior in-house MLOps stack.
  3. Container ID & damage OCR at terminal gates

    38 ms median latency at the gate — on the same Jetson the budget already approved.

    Maersk runs container-ID and damage-classification inference on edge hardware at terminal gates across Singapore. FGCV’s X1 runtime delivered the latency headroom the team needed to retire a separate TF Serving cluster; container ID reads dropped from a multi-stage pipeline to a single forward pass. Monitoring catches drift per gate so the ops team can re-route reviewers before errors compound.

    Result 38 ms median inference latency on Jetson Orin, benchmarked against TF Serving on equivalent hardware.
  4. Planogram & shelf compliance at store edge

    A versioned CV platform that store ops can actually trust.

    7-Eleven’s retail-CV team needed shelf and planogram-compliance models that could be retrained weekly across thousands of stores without an integration project each cycle. FGCV’s dataset versioning, evaluation gates, and edge runtime made each release a routine change — not a six-month integration project. Drift monitors flag stores whose camera exposure has shifted before compliance scores degrade.

    Result Weekly retraining cadence with zero dedicated integration engineering per release.

FGCV does not publish a logo wall. We publish the four named deployments we have full permission to detail, plus the aggregate benchmarks across all 2024 customers on the section below.

Quantified across 312 enterprise deployments in 2024.

Cross-customer proof points that aggregate the four deployments above — and the 308 behind them. Every number is sourced from the FGCV 2024 deployment review.

time-to-production 4.8× vs. in-house MLOps stacks (median across 312 deployments, 2024)
median inference latency 38 ms FGCV-X1 edge runtime, Jetson Orin & T4, vs. NVIDIA Triton and TF Serving
labeling cost reduction 71% median across manufacturing defect datasets using FGCV’s active learning loop
Fortune 500 in production 24 / 500 Named deployments include Siemens, BMW, Maersk, and 7-Eleven

Want the stack sidebar for your environment? Bring a sample dataset — we’ll show you the pipeline.

04 Deployment topology

What an FGCV production stack looks like from labeled dataset to monitored GPU inference.

Every case study below runs on the same four-stage topology. Read it once, then map the Siemens, BMW, Maersk, and 7-Eleven outcomes onto your own environment — data ingestion, training, edge runtime, and monitoring are the levers that determine your time-to-production.

Maintained by 47 full-time CV engineers from Tesla Autopilot, Google DeepMind, and Stanford SAIL. SOC 2 Type II, ISO 27001, and HIPAA certified — with documented on-prem and air-gapped deployments at three Fortune 100 manufacturers.

Data & labels Training & eval Edge runtime Monitoring
01

Labeled dataset

Multi-modal annotation across detection, segmentation, tracking, and OCR — with built-in active learning loops that cut labeling cost by 71% on manufacturing defect sets.

  • regions19 countries
  • locales11 languages
  • evalFGCV Eval Suite, 140+ citations
02

Training & versioning

Versioned datasets, model lineage, and reproducible runs — measured at 4.8x faster time-to-production across 312 deployments in 2024 versus in-house MLOps stacks.

  • deploys312 measured in 2024
  • speedup4.8x vs in-house
  • sdk14.8k GitHub stars
03

Edge runtime

FGCV-X1 edge runtime delivers 38 ms median inference latency on Jetson Orin and T4 hardware, benchmarked against NVIDIA Triton and TF Serving on equivalent targets.

  • latency38 ms median
  • targetsJetson Orin / T4
  • vsTriton, TF Serving
04

Monitored inference

Drift detection, data integrity checks, and reviewer queues — designed for teams that cannot afford a model drift incident or a missed defect in production.

  • retention92% gross / 148% net
  • nps74 enterprise (Q1 2025)
  • modescloud / on-prem / air-gap
Annotated architecture diagram of the FGCV four-stage production stack from labeled dataset to monitored GPU inference.
fig.04 — Four-stage FGCV topology, the same stack deployed at Siemens, BMW, Maersk, and 7-Eleven.
05 Next step

Get a 30-minute technical demo mapped to your stack, your region, and your latency targets.

Walk through your labeled dataset, target hardware, and inference budget with an FGCV solutions engineer. We'll come back with a written deployment plan, a reference architecture, and the measured latency you can expect on your own hardware — no slideware.

  • Stack review with a solutions engineer (not an SDR)
  • Region-specific deployment: cloud, on-prem, or air-gapped
  • Latency benchmark on your Jetson Orin or T4 target hardware
demo / 30 min Mon–Fri · 09:00–18:00 PT

Book a technical demo

Bring a model, a dataset, or just a hardware target. We'll show you the same four-stage topology running in your environment.

Solutions engineer
FGCV, Inc. · San Francisco
Direct line
+1 (415) 555-0188
Office
548 Market St, Suite 22417