3 of 12 are Testing & QA work, shown first
Testing and defect tracking
API upload testing
Real documents uploaded through the API in Postman.
PostmanREST APIs
Checked field by field
Every extracted field compared with the scanned source, by hand.
Reproducible defects
Each mismatch logged in Jira with the file, the field, the expected value and steps to reproduce.
Jira
Regression re-tests
The set rerun after every prompt or model change.
User stories
User stories written in Jira alongside the defects.
Choosing the right model: GPU benchmarks
Rented GPUs
GPU instances on Akamai Cloud and Google Cloud with NVIDIA L4 and RTX 6000 Pro.
Akamai CloudGoogle Cloud
Tested from my machine
Deployed each model and tunnelled ports to test it locally.
DockerSSH tunnels
Same documents for every model
5+ open-source OCR and vision-language models run on the same labelled set.
NuExtract 3PaddleOCRPaddleOCR-VLDoclingGOT-OCR
Scored by hand
Every field scored; checkboxes got their own benchmark because they were the hardest part.
Accuracy against cost
Field-level accuracy measured against cost per document.
Bedrock ruled out
AWS Bedrock Knowledge Bases tried for document Q&A; made-up answers on missing checkbox marks were the blocking issue.
AWS Bedrock
Troubleshooting a live platform, often overnight
Reproduce first
Run the same file through the same path until the failure shows up again.
Postman
Follow the record
Trace one record ID through each service's logs over SSH.
LinuxSSH
Network checks
Ports, IPs and firewall rules when a model endpoint stops answering after a change.
Payload and mapping bugs
Payload and field-mapping failures between services.
Written up clearly
Cause, fix and how it was verified, in Jira, and explained to the team in plain terms.
Jira