ML Pipeline & Backend Engineer — Python, LLM/VLM, spec-driven agents

About the project We're building Small Language Model (SLM), an AI system that turns construction drawings (PDFs) into structured steel takeoffs. It's a hybrid pipeline: a deterministic ML spine (object detection + geometry + engineering rules) augmented by gated LLM/VLM reasoning — not a thin wrapper over a chat model. You'll own the pipeline and AI layers, not the UI (frontend is covered). What you'll own - The extraction pipeline — merge raw ML detections with geometry, run validation/confidence gates, and post-process into clean structured output ("post-ML"). - Small, efficient models (SLMs) — page/element classifiers, embedding + linear-probe heads, and fine-tuning small models for on-device/low-cost inference. - Spec-driven agents & prompts — structured, reproducible LLM workflows (GitHub spec-kit–style) for a catalog of task agents (takeoff, estimator, QA), with structured-output contracts — not ad-hoc prompting. - Backend API services — improve and extend the Python (FastAPI) services behind the pipeline. - LLM/VLM integration — leverage Qwen (vision + text) served via vLLM, with guided/structured decoding. Must-have skills - Strong Python + backend/API engineering (FastAPI) - ML pipeline engineering — data → model → post-processing at production quality - LLM/VLM prompt & agent engineering, ideally spec-driven (spec-kit or similar) with structured outputs / tool use - Hands-on with open-weight models (Qwen-class) and serving via vLLM - Post-ML processing — turning raw detections into validated, structured data - SLM / small-model work — fine-tuning, embeddings, probes, quantization Nice to have - MCP (Model Context Protocol) tools/skills - ONNX + GPU model serving (RunPod or similar) - Document AI / CAD / structural-engineering domain - Local-first data (SQLite), quantized/on-device inference