Senior Python Engineer — LLM + Computer Vision Document Extraction
Budget: $15 – $25 USD
I run a production Python pipeline that extracts structured data from complex
technical PDF documents — detecting objects, tracing linear features, and
measuring areas and regions — using a mix of vision LLMs and computer vision. It
works and is deployed, but I need someone who can take real ownership and push
accuracy, robustness, and cost in directions I can't get to fast enough on my
own.
The pipeline is staged: render the PDF, read the document's reference key to
learn what to look for, detect every instance on each page with a spatial vision
model, measure each one (OpenCV contour/skeleton tracing, with an LLM agent
fallback), and hand off structured results.
What I need help with:
- Improving detection accuracy and recall on dense, messy, real-world documents
- Sharpening the CV measurement path (contour/area, skeleton/linear tracing,
scale detection) and reducing fallback rate
- Hardening the pipeline — partial-page failures, retries, timeouts, idempotency,
keep-best logic — so jobs don't silently degrade
- Driving down LLM cost and latency without losing quality
- Strengthening observability so regressions are caught in data, not by eye
Stack:
- Python (async/concurrent, staged pipeline architecture)
- Vision LLMs — Anthropic Claude + Gemini / Vertex AI spatial detection,
structured tool-use output
- OpenCV + PyMuPDF for rendering and measurement
- Axiom · Sentry · Langfuse for observability and cost tracing
You're a fit if you've shipped production LLM pipelines (vision and
structured-output, ideally multi-provider), are genuinely strong in Python, and
have real computer-vision experience — OpenCV, contour/skeleton work, image
geometry. The hard part is the extraction itself, not any particular subject
matter.
Please share examples of similar LLM and/or CV pipeline work in your bid —
especially anything where you improved accuracy on noisy real-world inputs.
technical PDF documents — detecting objects, tracing linear features, and
measuring areas and regions — using a mix of vision LLMs and computer vision. It
works and is deployed, but I need someone who can take real ownership and push
accuracy, robustness, and cost in directions I can't get to fast enough on my
own.
The pipeline is staged: render the PDF, read the document's reference key to
learn what to look for, detect every instance on each page with a spatial vision
model, measure each one (OpenCV contour/skeleton tracing, with an LLM agent
fallback), and hand off structured results.
What I need help with:
- Improving detection accuracy and recall on dense, messy, real-world documents
- Sharpening the CV measurement path (contour/area, skeleton/linear tracing,
scale detection) and reducing fallback rate
- Hardening the pipeline — partial-page failures, retries, timeouts, idempotency,
keep-best logic — so jobs don't silently degrade
- Driving down LLM cost and latency without losing quality
- Strengthening observability so regressions are caught in data, not by eye
Stack:
- Python (async/concurrent, staged pipeline architecture)
- Vision LLMs — Anthropic Claude + Gemini / Vertex AI spatial detection,
structured tool-use output
- OpenCV + PyMuPDF for rendering and measurement
- Axiom · Sentry · Langfuse for observability and cost tracing
You're a fit if you've shipped production LLM pipelines (vision and
structured-output, ideally multi-provider), are genuinely strong in Python, and
have real computer-vision experience — OpenCV, contour/skeleton work, image
geometry. The hard part is the extraction itself, not any particular subject
matter.
Please share examples of similar LLM and/or CV pipeline work in your bid —
especially anything where you improved accuracy on noisy real-world inputs.