Fercedferced
Back to projects

AI / Automation

DocuMind.

On-premise pipeline that processes PDFs with Qwen 2.5 7B running locally. A single Go binary extracts, analyzes and returns structured JSON — zero external APIs, zero data to third parties.

GoLLMllama.cppS3Queue
~/documind
$./documind start
[queue]watching input/ ...
[llm]qwen2.5-7b loaded (4.4GB)
[dash]http://localhost:9090
[queue]processing factura-0042.pdf
[llm]extracted 12 fields → output/
[queue]✓ done in 3.2s
$_
01 / Why they called us

They wanted to automate document reading without sending sensitive data to external APIs.

02 / The problem they had

Manual PDF review was slow, inconsistent and hard to integrate with internal systems.

03 / How we solved it

We designed a local pipeline that processes documents, extracts fields and returns JSON other systems can validate.

04 / Technical approach

Go, processing queues, local/S3-compatible storage and Qwen 2.5 7B through llama.cpp running on-premise.

05 / Impact

Private, repeatable processing that plugs into existing flows without exposing documents outside the client's infrastructure.

06 / How long it took

6 to 8 weeks for prototype, testing with real documents and hardening the pipeline.

07 / What we needed from the client

Provide representative PDFs and define which fields were correct or actionable.

08 / Process friction

Medium: AI needs iteration with real documents, but local deployment lowers legal and security friction.