Skip to content

Systems Works

Independent software and infrastructure work outside the computer-vision series.

Production VLM Serving

Problem. PaddleOCR-VL needed correct model and library versions, multi-GPU execution, and production deployment without Docker.

System. vLLM and Hugging Face served the model across eight GPUs on two bare-metal Debian servers, four GPUs per host. GitLab CI and Ansible delivered the stack under the CTO's no-Docker constraint.

Result. Production throughput was approximately two pages per second end to end: 30 pages in 14 seconds.

Lesson and next step. Reproducible version selection and provisioning were part of model serving, not surrounding tasks. No later iteration is recorded.

OCRmyPDF Remote PaddleOCR Plugin

Problem. PDF OCR needed to use a remote PaddleOCR model at scale.

System. A custom OCRmyPDF plugin calls an OpenAI-compatible endpoint served through vLLM. Model instantiation was moved out of the per-document path.

Result. PDF processing was twice as fast as the baseline.

Lesson and next step. Lifecycle cost can dominate an otherwise remote inference path. No further version is recorded.

ACL Audit Engine Migration

Problem. 17,000 lines of audit logic in ACL had to move to a maintainable Python CLI without changing outputs.

System. A dedicated audit engine separates data transformation from test execution and reproduces ACL semantics with pandas.

Result. Equivalent outputs were validated against the original, with a 10x speed improvement.

Lesson and next step. Behavioral equivalence was the migration boundary; speed alone would not establish correctness. No later phase is recorded.

Legacy Haskell and Jupyter Revival

Problem. A Haskell codebase implementing a Jupyter kernel depended on an unavailable license server and could no longer be run.

System. A Nix flake describes shells, packages, executables, and the matching Python environment. The source was changed to remove the license-server dependency.

Result. All software functions were restored.

Lesson and next step. Environment reconstruction and source repair had to be solved together. The revival is recorded as complete.

GitLab Issue Tracking

Problem. The team could not see issue throughput over time.

System. FastAPI provides the backend, Vue.js the interface, Celery asynchronous collection, MongoDB persistence, and the GitLab API source data.

Result. The application tracks issue throughput in real time.

Lesson and next step. The operational metric became a maintained data pipeline rather than a manual report. No further work is recorded.

Reproducible Open Notebook Deployment

Problem. A local-only Open Notebook installation needed a repeatable production environment.

System. Ansible installs and provisions bare-metal hosts; GitLab CI runs the delivery pipeline.

Result. The handoff produced two reproducible production VMs.

Lesson and next step. Reproduction at handoff was the deliverable. No later phase is recorded.

Small Tools and Runtime Experiments

  • EchoPulse-CPP — C++ and CUDA ESN inference.
  • HotStart — a measured negative result in model loading.
  • Alpaca — a compact vLLM operations CLI.
  • Blackhole — a small rsync interface.
  • Telecommit — commit generation from Git diffs.
  • Hugging-Hat — test-time latent computation modules.