Skip to content
00Forecast-grade computer vision

The operating system for modern forecast-grade computer vision — measured in 38 ms, not quarters.

FGCV replaces the brittle patchwork of model training scripts, label tools, and inference glue that most CV teams stitch together. One versioned platform takes a model from labeled dataset to monitored, GPU-efficient production in days — built by practitioners from Tesla Autopilot and Google DeepMind, deployed in production at 24 of the Fortune 500.

Get a Production Demo Read the Docs
  • 38 ms median edge latency
  • 4.8× faster time-to-production
  • SOC 2 Type II / ISO 27001 / HIPAA
01The pipeline

One platform, three pillars.

Training, deployment, and monitoring are first-class modules inside FGCV — not third-party glue. Each pillar ships with its own SDK, CLI, and review surface, and every artifact is versioned end-to-end.

  1. /01

    Training

    Versioned datasets, label queues with active learning, and reproducible training runs on shared GPU pools. The FGCV Eval Suite — cited in 140+ peer-reviewed CV papers — ships in the box.

    active learning cuts label cost
    71%
    label languages
    11
    eval suite citations
    140+ papers
  2. /02

    Deployment

    One artifact, every target. The FGCV-X1 runtime compiles a model to cloud GPUs, Jetson Orin, x86 edge boxes, and air-gapped on-prem — with deterministic latency budgets you can hand to procurement.

    median edge latency
    38 ms
    time-to-production
    4.8× faster
    deployments benchmarked
    312 / 2024
  3. /03

    Monitoring

    Drift, label integrity, and inference health in one panel. Alert routing hooks into PagerDuty, Slack, and OpsGenie — with a documented on-call runbook per deployment.

    customer NPS
    74
    gross retention FY24
    92%
    net retention FY24
    148%
02Architecture

From labeled dataset to monitored production — one versioned pipeline.

Every artifact on the left has a content hash, a lineage record, and a deployable mirror. There is no “copy this model onto the box” step — the runtime pulls the same signed blob that your training run produced.

labeled dataset active-learning queue
FGCV core // training + eval
FGCV-X1 runtime
cloud GPU jetson orin on-prem / air-gap
drift + health monitoring alert routing audit log
// fig 02.1 — every arrow is a versioned artifact with a content hash.
03Numbers a procurement-aware engineer can quote

The numbers that move the evaluation from “interesting” to “shortlist”.

  • 4.8× faster CV time-to-production vs. in-house MLOps stacks n=312 deployments, 2024
  • 71% labeling cost reduction via built-in active learning loops manufacturing defect datasets
  • 38 ms median inference latency on the FGCV-X1 edge runtime vs. Triton & TF Serving on equivalent Jetson Orin / T4
  • 74 customer NPS across enterprise accounts Q1 2025
04Patchwork vs. platform

Patchwork versus platform — what your team is currently maintaining.

The right column is not a sales pitch. It is the inventory of internal tools, scripts, and dashboards your CV team is currently maintaining to keep a model running in production.

typical in-house stack

patchwork — 11+ moving parts

  • Custom training scripts per model family (detection, segmentation, OCR).
  • Third-party label tool with CSV exports and re-import scripts.
  • Hand-tuned Docker images rebuilt on every CUDA bump.
  • Inference server glue (TF Serving, Triton, custom Flask) per edge box.
  • Drift detection in a Jupyter notebook that one engineer owns.
  • Alert routing wired through PagerDuty by a side project.
  • Six-month integration project every time a new camera ships.
time-to-production: quarters · drift incidents: recurring
FGCV

unified platform — one versioned surface

  • Training SDK covering detection, segmentation, tracking, OCR.
  • Built-in label queues with active learning (11 languages, 19 countries).
  • One signed artifact per model, hash-verified end-to-end.
  • FGCV-X1 runtime: cloud GPU, Jetson Orin, on-prem / air-gapped.
  • Drift, label integrity, and inference health in one panel.
  • Alert routing into PagerDuty, Slack, OpsGenie — with a runbook.
  • New camera or line? Same pipeline, re-targeted in days.
time-to-production: days · SOC 2 / ISO 27001 / HIPAA
05Trust

Trusted where CV cannot fail.

Certifications

  • SOC 2 Type II
  • ISO 27001
  • HIPAA
  • on-prem & air-gap

Documented deployments at three Fortune 100 manufacturers.

In production at

  • Siemens
  • BMW
  • Maersk
  • 7-Eleven

Powers 24 of the Fortune 500 in production.

Bench

47

full-time CV engineers and former researchers from Tesla Autopilot, Google DeepMind, and Stanford SAIL.

Open core

14.8k☆

GitHub stars on the FGCV SDK · 3.2M monthly PyPI downloads · public roadmap voted on by enterprise customers.

Ready to see the pipeline on your own data?

Get a Production Demo Review the security posture →