Messy documents in. Clean, structured data out.
I build document AI pipelines that read real-world scans the way a careful person would — finding every character, understanding table structure and checking the result. For Docutize.ai I built the entire backend that turns scanned laboratory reports into one clean table — 98%+ character accuracy and 99% column matching, at about three seconds a page.
- 98%+
- character accuracy after detection correction
- ~3 s
- per page, full extraction
- 1 table
- merged from multi-page reports
Problems I'm usually brought in to solve.
Data is trapped in scans
Results, invoices or forms arrive as scanned PDFs that people re-type by hand.
Generic OCR isn't accurate enough
One misread digit or shifted column silently corrupts your data.
Every source looks different
Layouts vary by sender, and tables run across pages with changing column names.
Volume makes speed matter
Character-level vision and OCR are slow until the pipeline is optimised.
What you get.
01
Ingestion API
Upload multi-page documents and split them into pages for processing.
02
Image enhancement
OpenCV clean-up so faint, skewed or noisy scans become readable.
03
Detection + OCR
Object detection (YOLO) to locate text, then OCR on the cleaned-up regions.
04
LLM structuring
Language models that understand the document and organise the data.
05
Validation & correction
Custom algorithms that catch and fix wrong detections and invalid values.
06
Speed optimisation
Cython, memory optimisation, multi-threading and multi-processing.
01
Discovery callWe talk through the problem, your data and constraints, and what success looks like. You get an honest view of what AI can and can't do here.
02
Architecture & planA clear design — components, models, data flow and how quality will be measured — with a scoped first version.
03
Build & evaluateWorking software early, tested against real examples from your domain rather than demos.
04
Ship & improveDeployed to production, monitored, and tuned for accuracy, latency and cost as real usage comes in.