My part
Built extraction, local Qwen2.5-7B inference through llama.cpp, structured JSON generation, schema validation, repair, source anchoring, retries, evaluation, and tamper-evident audit logs.
Local LLM pipeline / independent project
LLM Local PDF
A local Python pipeline that turns sensitive Argentine legal PDFs into structured, reviewable JSON without uploading documents to a SaaS model.
At a glance
- Period
- 2026
- Role
- Python extraction prototype development
- Team
- Independent engineering project
- Users
- People reviewing sensitive Argentine legal documents
- Status
- Working offline pipeline; public source; automated test sources
- ModelQwen2.5-7B
- Runtimellama.cpp + GGUF
- EvaluationField-level golden set
- Testsautomated test sources
Tradeoff
Keep documents and inference local while turning probabilistic model output into a structured proposal that can be measured, repaired, and reviewed.
Evidence
Public repository, documented local deployment, and automated test sources. The case makes no production-adoption or accuracy claim beyond the implemented evaluation workflow.
Related work · OficIA
A review shell for sensitive PDFs
OficIA is a separate product from this prototype. I designed and built its review interface: work queue, source and instructions, evidence per field and human approval.
Review shell: my work · Extraction pipeline: owned separately
Illustrative reconstruction with synthetic data, not connected to the real product. These are not original OficIA screens.
Source text
Synthetic notice. The requested party has 10 business days to submit the requested documents.
The highlighted text is the evidence for the proposed value.
Deadline
10 business daysProposed · check it against the source
Approval stays locked until a person confirms the field.
Confirm the field before approving.
01
Local inference is a product boundary
The pipeline extracts text and runs Qwen2.5-7B through llama.cpp on the local machine. Documents and inference do not leave the environment for a hosted model.
02
Model output must become a valid proposal
The LLM produces structured JSON, but the workflow does not trust the first response. JSON repair, schema validation, field-level source anchoring, and bounded retries either produce a reviewable result or an explicit failure.
03
Evaluation sources
The public repository includes a golden evaluation script, schema validation, source anchoring, review reports, and tests. These sources make the prototype available for inspection; no measured accuracy or production-adoption result is claimed here.
04
Reviewability over autonomy
The output is a proposal for human review, not a legal decision or automatic external action. Tamper-evident audit logs preserve what ran and how the result changed.
05
Related work: OficIA
OficIA is a separate product for reviewing sensitive PDFs. I designed and implemented its review shell: the work queue, the source and instruction workbench, field-level evidence, and human approval before publication. Its core extraction pipeline is owned separately.
Project linksGitHub