Case Studies

Technology proven in production at some of the world’s leading organizations.

Request demo

Technology developed and deployed with the world’s leading organizations

Google

Google used Snorkel to replace 100K+ hand-annotated labels in critical ML pipelines.

Problem

Content, product, and event classification problems change too fast to hand-label, even with significant annotation budget.

Solution

We deployed early versions of Snorkel Flow's core technology with three high-impact teams at Google, repurposing many organizational resources as labeling functions.

Results

Hours of labeling function development replaced 10-100K+ hand labels, significantly impacting the bottom line and acceleration of ML solution adoption.

6 MONTHS

of hand-labeling data replaced in 30 mins

<0hrs

To develop the first custom ML model

52%

performance improvement

+0%

Accuracy for contract classification

100K

hand labels replaced with programmatic approach

Contracts processed in minutes

Top U.S. Bank

A top U.S. bank uses Snorkel Flow to quickly build AI applications that classify and extract information from their documents.

Problem

The bank estimated that, for a time-sensitive use case, hand-labeling data would take over a month.

Solution

With Snorkel Flow, the team produced a solution that was over 99% accurate in under 24 hours.

Results

The resulting AI application could be quickly and easily adapted to new problems and business lines.

99 .1%

Snorkel Flow Accuracy

<0hrs

To develop the first custom ML model

<24hrs

From problem start

+0%

Accuracy for contract classification

>250K

# Documents processed

Contracts processed in minutes

Apple

Apple built applications with an internal Snorkel-based system that answered billions of queries in multiple languages and processed trillions of records with up to 2.9x fewer errors.

Problem

Apple needed a system that supported engineers facing contradictory or incomplete supervision data.

Solution

Apple built a solution called Overton which utilized Snorkel’s framework of weak supervision to overcome cost, privacy, and cold-start issues.

Results

Overton achieved a 12%+ bump in F1 score by going from 30K to 1M data labels.

12%

bump in F1 score

<0hrs

To develop the first custom ML model

2.9x

fewer errors with Snorkel-based applications

+0%

Accuracy for contract classification

32x

more labels generated

Contracts processed in minutes

Intel

Intel used Snorkel to replace a high-cost, high-latency crowdsourcing pipeline and accelerate sales and marketing agents.

Problem

Rapidly changing sales goals make social media monitoring difficult to maintain.

Solution

Deployed a proto version of Snorkel (Snorkel Osprey) to replace months-long crowdworker labels with cheap & fast programmatic labeling.

Results

Better performance and major cost savings in Sales & Marketing and Advanced Analytics.

6 MONTHS

of crowdworker labels replaced

<0hrs

To develop the first custom ML model

+18.5

performance improvement

+0%

Accuracy for contract classification

+28.5

Coverage percentage points

Contracts processed in minutes

Fortune 50 Bank

A Fortune 50 bank achieved a 25+ point performance gain over a black box vendor solution for news analytics application with Snorkel Flow- in just a few weeks.

Problem

The bank needed an accurate way to tag companies in unstructured news text, link them to identifiers (e.g., stock tickers), and classify mentions by sentiment and other aspects.

Solution

The bank used Snorkel Flow to develop an AI-powered news analytics application that monitors target companies' press coverage in unstructured data feeds.

Results

With Snorkel Flow, the team achieved a 25+ point performance gain over a legacy vendor system and internal heuristic approaches.

45x

faster compared to hand-labeling

<0hrs

To develop the first custom ML model

+90

F1 score for news analytics application

+0%

Accuracy for contract classification

+25%

point performance gain over black box vendor system

Contracts processed in minutes

Stanford Medicine

Researchers at Stanford Medicine used Snorkel to label medical imaging & monitoring datasets, replacing person-years of hand labeling with several hours of using Snorkel.

Problem

Labeling training data for triaging models takes person-months to person-years of radiologist time.

Solution

We deployed a cross-modal Snorkel pipeline, matching or exceeding the performance of painstakingly gathered manual labels in hours.

Results

Currently being tested for deployment in Stanford & Department of Veteran Affairs (VA) hospital systems.

Person-months of labeling replaced

<0hrs

To develop the first custom ML model

94%

ROC AUC Performance

+0%

Accuracy for contract classification

50K+

Images labeled in minutes

Contracts processed in minutes

Fortune 50 Bank

A Fortune 50 bank extracts financial information from PDFs with 99% accuracy in milliseconds using a financial spreading application built with Snorkel Flow.

Problem

The bank needed to extract structured financial data from balance sheets and income statements (hOCR PDF) from private company financials.

Solution

The bank used Snorkel Flow to develop an AI-powered financial spreading application that parses textual and spatial/visual data features.

Results

With Snorkel Flow, the team achieved superior performance with greater generalizability (2x coverage) compared to a purely rules-based approach.

coverage compared to rules-based approach

<0hrs

To develop the first custom ML model

99%

extraction accuracy

+0%

Accuracy for contract classification

45x

faster compared to hand-labeling

Contracts processed in minutes

Tide

A UK based fintech company, used Snorkel to match receivable invoices from the mobile app with incoming transactions.

Problem

Tide needed to label matching invoices with transactions that required investing highly paid subject matter experts’ time in hand-labeling historical data.

Solution

Used Snorkel to programmatically label data, extract information, and harness business knowledge by creating labeling functions.

Results

Achieved 97.6% accuracy to detect transactions made for a particular invoice. Created training data programmatically replacing 1000 hours of hand labeling.

Days to create training dataset& deploy model

<0hrs

To develop the first custom ML model

97%

ML model accuracy

+0%

Accuracy for contract classification

Invoices processed

Contracts processed in minutes