OCR Toolbox
Turns photos of handwritten HSE forms into structured data.
Visit live ↗Demo requires credentials
Problem solved
Handwritten HSE forms must be re-entered manually: slow, costly, and error-prone.
My role
Designer of the processing pipeline and developer of the backend and review interface.
Stack & architecture
- FastAPI
- Python
- OpenCV
- RapidOCR
- Vision LLM
Notable technical decisions
- Pipeline: OpenCV preprocessing → form-type classification with local RapidOCR → content extraction via a vision LLM (one call per document) → ISO date normalization + fuzzy-matching names to the employee master → manual browser review → processor.
- Form types are defined as templates (field + type + keywords) so new forms require no code changes.
- The built-in toolbox_kpi processor computes Toolbox Meeting attendance per 7-day period and fills the KPI CPM.xlsx template without breaking its formulas.
Operational impact
Field HSE form data becomes structured and analysis-ready without manual re-entry.