Medium

Aditya Mangal

Aditya Mangal

·

Sep 30

I Found 44 PHI Leaks in a “Secure” Document Pipeline — Here’s the 12-Point Audit

Ask an engineering team where protected health information lives in their system and you will usually get a confident answer: the database…

I Found 44 PHI Leaks in a “Secure” Document Pipeline — Here’s the 12-Point Audit

Aditya Mangal

·

Sep 4

I Tried Extracting Tables from PDFs in C# — Here’s Where It Breaks

By Aditya Mangal — Senior AI Engineer. Every result here was produced on my own Windows laptop, .NET 10, no cloud services.

I Tried Extracting Tables from PDFs in C# — Here’s Where It Breaks

Aditya Mangal

·

Sep 1

I Tested 4 Ways to OCR Images in C# — One Took Just 41 ms

By Aditya Mangal — Senior AI Engineer. Every number here was measured first-hand on my own Windows laptop. No cloud APIs, no Python — just…

I Tested 4 Ways to OCR Images in C# — One Took Just 41 ms

Aditya Mangal

·

Aug 24

Extract Text from PDF in C# Without Cloud APIs: Detect the Text Layer Before You OCR

By Aditya Mangal — Senior AI Engineer. Every result here was produced on my own Windows laptop, .NET 10, no cloud services.

Extract Text from PDF in C# Without Cloud APIs: Detect the Text Layer Before You OCR

Aditya Mangal

·

Aug 12

The Ultimate OCR Benchmark: 15 OCR Systems Tested on My RTX 4070 Laptop

Classic engines, OCR-specialist VLMs, generalist VLMs, and full document pipelines — same invoice, same RTX 4070, real code, real output…

The Ultimate OCR Benchmark: 15 OCR Systems Tested on My RTX 4070 Laptop

Aditya Mangal

·

Jul 30

Surya-2 vs GOT-OCR 2.0: Two ~600M OCR VLMs, One 8 GB Laptop, Measured

They’re almost the same size — 650M and 561M — and both are small open OCR vision-language models you can run locally. But they make…

Surya-2 vs GOT-OCR 2.0: Two ~600M OCR VLMs, One 8 GB Laptop, Measured

Aditya Mangal

·

Jul 28

I Ran Surya-2 on My RTX 4070 Laptop: Can a Modern OCR VLM Really Run on an 8GB GPU?

Surya-2 is a 650M-class vision-language OCR model that emits structured, typed, coordinate-tagged output. The internet says “pip install…

I Ran Surya-2 on My RTX 4070 Laptop: Can a Modern OCR VLM Really Run on an 8GB GPU?

Aditya Mangal

·

Jul 22

PaddleOCR-VL with vLLM: A Complete Performance Benchmark (76× Faster)

I Ran PaddleOCR-VL on vLLM — and the 273s/page Became 3.6s

PaddleOCR-VL with vLLM: A Complete Performance Benchmark (76× Faster)

Aditya Mangal

·

Jul 21

I Tried Running MonkeyOCR Locally — Here’s Where an 8GB Laptop Hits Its Limit

MonkeyOCR is one of the most capable document-parsing systems of 2026 — a three-stage Structure → Recognition → Relation pipeline. The…

I Tried Running MonkeyOCR Locally — Here’s Where an 8GB Laptop Hits Its Limit

Aditya Mangal

·

Jul 20

The Hidden Cost of Local OCR: dots.ocr vs PaddleOCR-VL on an 8GB Laptop

Two of the best open OCR vision-language models, both meant to run locally, on my RTX 4070. I expected a clean accuracy-and-speed race…

The Hidden Cost of Local OCR: dots.ocr vs PaddleOCR-VL on an 8GB Laptop

Aditya Mangal

Aditya Mangal

AI Systems Architect | Building production-grade AI agents & RAG pipelines | Writing about real-world AI engineering, not just hype.

Following

See all (3,406)

Help

Status

About

Careers

Press

Blog

Store

Privacy

Rules

Terms

Text to speech