Shipped · live demos by appointment

Power OCR: private OCR for scanned PDFs, on your own hardware

Power OCR turns scans, faxes and image-only PDFs into searchable text, tables and searchable PDFs. Staff can then ask questions and get answers that cite the page. Every file is virus-scanned first, and nothing is kept unless you ask.

Power OCR is designed to run locally: on your own machine, or in a cloud space you control. It does not run through third-party AI companies. It is built for teams with PII or HIPAA obligations, such as law firms, clinics, insurers and records offices. Power OCR is not HIPAA certified and makes no compliance claim. Once installed, it keeps your documents on systems you control.

A live demo runs on our RTX 5090 workstation during a booked session. You can also try a pilot on your own sample documents, under an NDA, with a deletion record for each job.

121

pages per minute, end to end, measured on our private AI test server (RTX 5090)

99.4–100%

characters read correctly on six new test documents, same server

0

connections to outside hosts in our monitored run

MIT

licence of the OCR model; the engine is Apache-2.0

How demo mode works

  • Demo mode is a switch. We turn it on for a client’s test, and off again when the test ends. While it is off, as it is now, the app is visible but none of its services run.
  • Your files are held only for the test. During a demo, your files live on our server only while the test runs: in GPU memory, and as temporary working files (uploads, exports and the search index). When the demo ends, we delete all of it and unload the model from the GPU. You can also delete a job sooner. Each job gets a deletion record listing the fingerprint (SHA-256) of each file received and everything deleted, and you can have a copy. It is a normal deletion, not a forensic disk erase. We collect no more data than we need to deliver the product.
  • Search takes seconds. Once your files are loaded and processed, searching across all of them takes seconds, and each answer points to the page it came from.
  • After the demo, it runs on your side. Power OCR is designed to run locally, on your own machine or in a cloud space you control. It does not run through third-party AI companies.

The Power OCR app

Demo mode: OFF

Services are switched off. You can see every part of the app, but nothing here uploads or reads a file. The OCR engine only runs during a booked live demo on our RTX 5090 workstation.

Request a live demo

Services

Scan to text

Scans, faxes and photos of pages become Markdown and plain text, with headings and tables kept.

Searchable PDF

The same pages with a hidden text layer, ready for any PDF reader or records system.

Ask your documents

Ask a question or press Summarize. Answers quote the page they came from.

Backlog conversion

Thousands of pages in one queue, with a time estimate before you start.

Malware scan first

ClamAV scans every upload before anything opens it. Infected files are blocked and never read.

Deletion record

Delete a document or a whole job when you ask. Its files, database rows and search entries are removed, and you get a record of what was deleted.

Workbench

Drag PDFs or images here (off in this preview)

Exports: Markdown, text, JSON, searchable PDF

Results

Finished documents appear here during a live demo.

Ask or summarize

Answers use only the documents uploaded in that session, and cite the page.

What the same app does in a live demo

The Power OCR workbench during a live demo: an invoice table and a degraded fax read in 2.1 seconds, with Markdown, text, JSON and PDF downloads and a Delete this document now button.
Two test documents read on the RTX 5090 in 2.1 seconds. Each has downloads and a delete button.
Ask or summarize: the question What is the total due? answered with Total due: $9,243.00 and a source citation to the invoice, page 1.
Ask quotes the answer and cites the page. These are our own generated test documents.

How long will my job take?

This estimator runs in your browser and sends nothing anywhere. It uses our measured local rates: 121 pages a minute on a typical job, 78.4 on dense real scans, and up to about 107 seconds to start the engine cold. The retyping row assumes average typing speed (51.56 words a minute) and 400 words a page, with no proofreading.

Power OCR, typical (one RTX 5090)about 6 minutes
Power OCR, dense real scansabout 8 minutes
One person retyping (8-hour days)about 8.1 working days

These are estimates, not quotes. Real time depends on page quality, layout and the hardware used. Paper still has to be scanned first, whichever OCR you use.

Compared with retyping and desktop OCR

Estimates for a typical office job. Power OCR times are for one RTX 5090 with the engine warm; a cold start adds about 1.5–2 minutes. Retyping assumes one person at average speed. Acrobat OCR is machine time on one PC.

JobPower OCRRetyping by handAdobe Acrobat OCR
10-page contractabout 7 sabout 1.3 hoursabout 11–41 s
500-page case fileabout 4–7 minabout 64 hours (8 working days)about 9–34 min
10,000-page backlogabout 1.4–2.1 hoursabout 1,280 hoursabout 3–11 hours
100,000 pagesabout 14–21 hoursabout 12,800 hours (6.4 work-years)about 30–114 hours

Sources: typing speed from Dhakal et al., “Observations on Typing from 136 Million Keystrokes”, CHI 2018 (doi). Acrobat speed from Adobe’s ClearScan post (1.09 s a page at 300 dpi, 2009) and a 2026 vendor comparison (about 4.1 s a page). Power OCR figures are our own runs on Nov 11, 12 and 14, 2026; runs over 10,000 pages are extrapolated.

Example build: one RTX 5090 workstation

This is the build we measured. An HTTPS proxy is the only open port. The app, the ClamAV scanner and the OCR engine sit on a private network with no route to the internet. Uploads are scanned first, and if the scanner is down, uploads are refused. The audit log records who did what and when. It never holds document text.

Power OCR local architecture: staff reach an HTTPS proxy; behind it the app, ClamAV scanner and OCR engine run on a private network with no internet route, on one GPU server.
What lives where on your own GPU server.
One 120-page document through Power OCR with measured stage times: virus scan, page images, OCR on the GPU, exports and index.
One 120-page document, with the time each step took. Reading the pages on the GPU is about 85% of the time.
Where Power OCR can run: your own GPU server, your cloud account, a dedicated rented server, managed by AI Playground, or a pilot on our workstation.
Where Power OCR can run.
Pilot on AI Playground's workstation: your documents travel over a private tunnel, are read, returned, and deleted when the job ends.
Pilot on our workstation over a private tunnel. Nothing is kept, and you get a deletion record.

Our private server or AWS: speed or scale

AWS gives you scale over speed. Our private AI test server is faster per job but handles smaller data loads. AWS can run many GPUs at once for very large backlogs, in a cloud account you control.

Private AI test server (RTX 5090)AWS, one NVIDIA L4 GPU (planned)
GPU memory32 GB GDDR7, 1,792 GB/s (NVIDIA spec)24 GB GDDR6, 300 GB/s (NVIDIA spec)
Speed, one GPUMeasured: 121 pages a minute end to end; 78.4 on dense real scansNot measured. Expected to be roughly 4–6 times slower per GPU, estimated from the specs
ScaleOne GPU, one job at a timeAdd GPU hosts in parallel for large loads; GPUs stop when idle

Your own AWS account Planned, not yet run

The same design for a client’s own AWS account: uploads land in an encrypted quarantine bucket, are scanned, then read on one NVIDIA L4 GPU host. A budget alarm turns the GPU off at the limit. This version is designed but has not been run yet.

Planned AWS design: web firewall and load balancer, sign-in, S3 quarantine and clean buckets, ClamAV scanner, one L4 GPU host, budget alarm and kill switch, audit trail with no document text.
AWS: what lives where (planned).
Planned AWS flow for one document: upload, quarantine, scan, read on the GPU, results, auto-delete.
AWS: one document, step by step (planned).

Download the Power OCR design overview (PDF) for IT, security and operations buyers: measured timings, costs, hosting options and the questions to answer before an install.

Why your documents stay yours

  • Before using the open-source OCR model, we audited its code and dependencies. We removed or switched off 17 risky pieces, including a remote-code auto-update, library telemetry, eval() on model output, unsafe file loading, and document text written to logs.
  • On Nov 14, 2026 we ran six new documents with a packet monitor inside each container. The app, engine, scanner and proxy made 0 connection attempts to outside hosts. The only non-local name lookup was the engine asking the private resolver about its own 0.0.0.0 address at start-up. It carries no document data.
  • Every model file and container is pinned by its SHA-256 hash. The model weights match the publisher’s published hashes.
  • We can sign an NDA, and every job comes with a deletion record. In our test on Nov 14, 2026, ending the demo returned the data folder and database to their exact pre-test file and row counts, with no trace of the test text left. It is a normal deletion, not a forensic disk erase.

The model behind it

Power OCR runs Unlimited OCR 2, an open-source vision-language OCR model of about 3.3 billion parameters, published by Baidu under the MIT licence. It builds on DeepSeek-OCR (MIT). It runs locally on the SGLang engine (Apache-2.0) and uses about 10 GB of GPU memory in our demo setting. Ask and Summarize quote your documents directly, with no chatbot involved. If you want summaries written by an AI, you can plug in your own local model. There is no AGPL code in the build: we removed PyMuPDF from the engine on Nov 14, 2026, and PDFs are rendered with pypdfium2 (Apache-2.0 / BSD). A licence inventory of every package is part of each install.

Services: start with a free consultation

Every engagement starts with a free consultation. After that, choose a one-time setup that you own outright, or a lower setup rate plus a monthly retainer that keeps us on call for AI/ML tech support for Power OCR.

Step 1

Free consultation

Free

We talk through your documents, your volumes, your PII or HIPAA obligations, and where Power OCR has to run.

  • A look at a few sample document types and how many pages you handle
  • Your own hardware or a cloud space you control: which fits, and what it needs
  • Optional live demo on our private AI test server, with demo mode switched on only for your test and your files deleted when it ends
  • A written recommendation: one-time setup or support retainer

Option A

One-time setup

Quoted after your consultation

We install and configure Power OCR on your hardware or in your own cloud account, train your team, and hand it over. You own the install.

  • Install and configure on your own GPU machine, or in a cloud account you control
  • Virus scanning, exports, the searchable library if you want it, and deletion records set up
  • A test run on your sample documents, with the measured speed and accuracy written up
  • Training for your staff and IT, plus a written runbook

Option B

Support retainer

Lower setup rate plus a monthly retainer, quoted after your consultation

The same setup at a lower rate, plus a recurring retainer that keeps us on call for AI/ML tech support for Power OCR.

  • On call for fixes, with remote help when something breaks
  • Model and engine updates and security patches, tested before they reach you
  • Accuracy tuning when you add new document types
  • Help scaling up: more GPUs, more users, or a move to your own cloud

Third-party software: Power OCR includes ClamAV (GPL-2.0), run unmodified as a separate component, and uses NVIDIA CUDA libraries under NVIDIA’s own licence terms. For client installs, on your own hardware or in your own cloud, we build Power OCR on your machine, and you accept those third-party terms.

Book a free consultation

Prices depend on your documents, volumes and hosting, so we quote them after the consultation. No package includes a HIPAA certification or compliance claim.

Questions and answers

What is Power OCR?
Power OCR is private OCR software for offices. It turns scanned PDFs, faxes and photos of pages into searchable text, tables and searchable PDFs. Your staff can then search the documents or ask questions, and each answer cites the file and page it came from. It runs on one NVIDIA GPU in your office, or in your own cloud account.
Can I use it on this website?
Not yet. The app on this page is shown with its services switched off. Demo mode is a switch: we turn it on for a client's test on our private AI test server (RTX 5090), then off again. During a demo, your files are held on our server only for the test, in GPU memory and as temporary working files. When the demo ends, all of it is deleted and the model is unloaded. Once your files are processed, searching across all of them takes seconds. We collect no more data than we need to deliver the product. Email us to book one.
Is Power OCR for PII and HIPAA data?
It is built for teams with PII or HIPAA obligations. Power OCR is designed to run locally, on your own machine or in a cloud space you control, and it does not run through third-party AI companies. It is not HIPAA certified, and we make no compliance claim. Your own HIPAA program, agreements and controls still apply.
How fast is it?
On our private AI test server (one RTX 5090) we measured 121 pages a minute end to end on a 120-page job, and 78.4 pages a minute on 782 dense real court scans run as one document. The engine takes about 1.5 to 2 minutes to start from cold. A 500-page file takes roughly 4 to 7 minutes once the engine is warm. These are estimates from our runs, not a quote.
How accurate is it?
On six new test documents (a contract, an invoice table, a two-column page, a degraded fax, a typewriter memo and fine print), 99.4% to 100% of characters were read correctly. Very small print on dense product-label exhibits can be missed. We score your own sample documents during a pilot.
Does it send my documents anywhere?
No. We removed or switched off 17 risky pieces of the open-source original, including a remote-code auto-update, library telemetry and eval() calls on model output. In a monitored run on Nov 14, 2026, the app, OCR engine, virus scanner and proxy made 0 connection attempts to outside hosts.
Which AI model does it use?
Unlimited OCR 2, an open-source OCR model of about 3.3 billion parameters published by Baidu under the MIT licence. It is built on DeepSeek-OCR, which is also MIT. It runs locally on the SGLang engine (Apache-2.0) and uses about 10 GB of GPU memory. Ask and Summarize quote your documents directly. No chatbot is involved unless you add your own model. There is no AGPL code in the build; PDFs are rendered with pypdfium2.
How do I know my demo files were deleted?
When the demo ends, we delete every demo job's uploads, exports, database rows and search entries, and unload the model so its GPU memory is released. Each job gets a hash-chained deletion record listing the SHA-256 of each file received, everything deleted, and GPU memory before and after. It is a normal deletion, not a forensic disk erase.
How do I get Power OCR set up, and what does support cost?
Start with a free consultation. Then choose a one-time setup on your hardware or in your own cloud, with training, or a lower setup rate plus a monthly retainer that keeps us on call for AI/ML tech support. Prices are quoted after the consultation. Power OCR includes ClamAV (GPL-2.0, run unmodified as a separate component) and uses NVIDIA CUDA libraries under NVIDIA's own licence terms. For client installs we build it on your machine, and you accept those third-party terms.
We have 10,000 pages or more. Do you connect to where our files already live?
Yes. For large jobs, Power OCR connects read-only to your existing storage, such as SharePoint, an S3 bucket or a network share, or watches a drop folder you choose. Your originals stay where they are. Power OCR reads them and writes only its results.
What hardware do I need?
An NVIDIA GPU with at least 12 GB of memory (16 GB recommended), plus Docker. Our example build is one RTX 5090 workstation. A cloud version for your own AWS account is designed but has not been run yet. AWS gives scale over speed: we expect one AWS L4 GPU to be several times slower than our RTX 5090 (an estimate from the spec sheets, not a measurement), but you can run many at once.

More from AI Playground

Need a different private AI tool built for your office? See Hire Me. Already have a diagram stuck in a scanned PDF? Try Diagram Tool.