I build automation for work that people are still doing by hand.
Most of it starts the same way: someone is copying numbers between spreadsheets, re-typing text off scanned documents, or running the same five steps every morning. I turn that into something that runs on its own, and I keep it small enough that one person can maintain it.
What I work on
| Document and image pipelines | OCR at scale, table extraction, batch conversion. Local-first, so nothing leaves the machine. |
| LLM systems that run in-house | Retrieval over private document sets, fine-tuned small models, local inference. For teams that cannot send their data to an API. |
| Computer vision for physical operations | Object detection and tracking fine-tuned on the customer's own footage, not on a generic dataset. |
| Internal tools | FastAPI and Next.js apps, containerized, deployable on a single box. |
Selected work
| Repository | What it does |
|---|---|
| ocr-smart-tools | Local OCR toolbox built on PaddleOCR. Batch processing, layout preservation, optional LLM cleanup pass. Runs fully offline. |
| paddleocr-vl-setup | Deployment kit for PaddleOCR-VL: four working demos, install scripts, English and Chinese docs. |
| network-debug-notes | Four production failures traced to root cause, including the wrong theories I chased first. |
How I work
Two habits that have earned their keep:
I do not trust a theory that explains everything. A story that accounts for every symptom is usually the one that costs the most time. I isolate variables physically before I reason about them, because reasoning is where the wrong story gets in.
I verify after changing, not before. Silent failures cost more than loud ones. A config edit that did not apply, a script that matched nothing and exited zero: these are the expensive bugs, and they only show up if you read the state back.
The network-debug-notes repo is the honest version of this, including four cases where I was confidently wrong before I was right.
Stack
Python · FastAPI · PyTorch · Ultralytics · PaddleOCR · Next.js · Docker · ffmpeg
Get in touch
Open to consulting work on document automation, private LLM deployment, and applied computer vision.
For consulting via 写春秋管理咨询, visit the site. Personal notes and updates: @odpanda on X. Or open an issue on any repository above—describe the manual process you want removed and I will tell you honestly whether it is worth automating.