Full-day live workshop · Recordings included

Choose, evaluate, and deploy OCR for PDFs.

Compare OCR models on real PDFs, build an evaluation harness, and deploy a production system.

Loan estimate closing-cost tables — Source PDF
**2 months**

We work through real PDFs, including selected participant submissions.

We’ll compare source pages with raw model outputs, find what failed, and decide what to change in the model, tests, or deployment.

Failure analysis

Find failures that look correct at first

OCR can complete text it cannot see, move a value into the wrong header, or preserve words while destroying structure.

Output format

Keep each value in the right row and column

OCR may extract every number but put it under the wrong heading. Your application then uses the wrong value.

Source tableIndividual income taxes | Total: 4,085.5

Product-ready record{"label":"Individual income taxes","total":"4,085.5"}

Product evaluation

Check the fields your application depends on

OCR may extract 0.3 and 3.0 but drop the temperature range they apply to. Your application can no longer interpret those values correctly.

Datasheet rowV_OS | LM358B | T_A = 0°C to +85°C | 0.3 | 3.0 | mV

Extracted record{"test_condition":null,"typ":"0.3","max":"3.0","unit":"mV"}

See all 15 public examples →

After registration, submit a PDF and describe what your product needs from it. We cannot review every submission live, so we’ll select participant PDFs for walkthroughs. That keeps the day grounded in the documents people are actually trying to put into production.

Do not submit confidential, regulated, or private material unless you have permission to share it. We will confirm how a selected example may be shown before using it with the cohort.

Curriculum

Choose the OCR model or VLM

Define whether the product needs text, Markdown, tables, layout, bounding boxes, or structured data
Compare managed APIs, traditional OCR pipelines, and open VLMs
Test quality, throughput, and cost on pages your product will actually receive

Compare the output with the PDF

Compare literal text, structure, tables, and regions against the original page
Separate model failures from rendering, preprocessing, and post-processing problems
Group repeated failures, then turn them into tests

Evaluate while you build

Use formative evals while building and summative evals before release
Connect OCR checks to downstream tasks such as extraction, search, and question answering
Build a regression set from real failures and trace downstream errors back to the document-processing stage

Put it into production

Trace a PDF from upload and storage through page rendering, queues, inference, validation, and delivery
Compare vendor APIs with self-hosting on Modal using throughput and cost estimates
Add retries and observability, then decide which parts of the system to own

Live walkthroughs

Compare OCR outputs. We will run the same pages through several tools, save the raw outputs, and compare where each tool fails.
Build the evaluation harness. We will create a small test set, mark failures on the source pages, compare configurations, and choose one to deploy.
Office hours and Q&A. Ask questions throughout the day and bring questions from your own workflow to the final session.

Build and evaluate an OCR system on real PDFs.November 14 · Full-day live workshop · Recordings included

“Practical tips and good examples in repos that helped make the recommendations more tangible.”
Robin TullyCTO, Forecast.ingFeedback from Isaac’s previous live course

Instructors

Isaac Flath

AI products · Education and research

Isaac has spent the last 10 years building AI products and has taught more than 4,000 students in courses on retrieval, evaluation, AI-assisted coding, and production AI systems. His work focuses on document processing, evals, traces, and the product workflows around them.

isaacflath.com

Joe Barrow

Senior Research Scientist · Adobe Research

Joe trains and serves vision-language models for document tasks. He previously led the machine learning team at Pattern Data and earned his PhD from the University of Maryland.

jbarrow.ai

Isaac’s students have included engineers and technical leaders from:

Enrollment

30 days of follow-up support included

Weekly office hours and unlimited async questions as you apply the workshop to your own documents and system.

“Instructors are highly available for questions and advice and provide a ton of additional tutorials and walkthroughs.”
Mike PowersAI EngineerFeedback from Isaac’s previous live course

Can I submit my own PDF?

Yes. Tell us what your product needs from it, and we may use it in a live walkthrough. We’ll confirm permission before showing a selected submission. We cannot promise that every PDF will be reviewed live.

Do I need to code?

No. The lectures, examples, and Q&A do not require you to write or run code. The notebooks and code walkthroughs are optional: use them yourself, share them with an engineer, or give them to a coding agent when you’re ready to implement.

What happens if I cannot attend the full day?

Every session is recorded and the recordings are available after the workshop. You can return to any session you miss. You’ll also receive the notebooks and 30 days of follow-up support.

$500

Live on November 14, 2026
Live instruction and Q&A
Working OCR comparison notebook
Evaluation harness notebook
Option to submit your own PDF
Recordings and course materials available afterward
30 days of follow-up office hours and async support