Machine Learning Engineer Geometry & Document Intelligence

We're a founder-led, early-stage company applying AI to a large, technical industry. We're in stealth, so specifics come later in the process — but the core of what we do is turning dense, highly structured technical documents into precise, quantitative, machine-usable data. The founding team comes from autonomous systems and robotics. Full details on product, market, and traction are shared under NDA during interviews. The Role About Us We're a founder-led, early-stage company applying AI to a large, technical industry. We're in stealth, so specifics come later in the process — but the core of what we do is turning dense, highly structured technical documents into precise, quantitative, machine-usable data. The founding team comes from autonomous systems and robotics. Full details on product, market, and traction are shared under NDA during interviews. The Role You’ll be the founding engineer on our core extraction pipeline. The problem spans deterministic computational geometry, document layout analysis, text understanding, targeted deep learning, and relational reasoning — and much of the job is judgment about which tool belongs where. Our philosophy: use learned models only where deterministic methods fail, and concentrate ML exactly where user feedback accumulates so the system improves continuously. What You'll Work On Vector geometry extraction. Parse vector documents into their primitive geometry — paths, lines, arcs, positioned text — using tools like PyMuPDF (fitz) and pdfplumber for PDFs and ezdxf for CAD interchange formats (DXF/DWG), and recover higher-level structures deterministically from geometric signatures, without relying on pixel-level inference. Layout and region analysis. Classify document regions (drawing areas, tables, notes, reference blocks) so downstream extraction operates on the right content — combining rule-based methods with lightweight learned classifiers where needed. Text extraction and spatial association. Extract embedded and OCR'd text — including rotated, multi-angle annotations, using engines like PaddleOCR or Google Document AI where OCR is required — and associate each annotation with the geometric element it describes, via proximity reasoning, leader-line following, and cross-referencing against tabular data elsewhere in the document. Object detection and segmentation. Train and deploy detection models (YOLOv8/v11 via Ultralytics or similar) for compact symbols that geometry alone can't disambiguate, and instance segmentation (Detectron2 / Mask R‑CNN class methods) where the deliverable is a polygon rather than a bounding box. Build the data engine: auto‑generating labels by rasterizing vector sources against their own known geometry, plus harvesting labels from in‑product user corrections. Relational / graph reconstruction. Go beyond detecting individual elements: build the geometric and rule‑based reasoning layer that reconstructs relationships between elements — adjacency, spanning, bearing, cross‑references — into a coherent graph from which precise quantities can be computed and audited. LLM/VLM integration. Use large multimodal models (Claude, GPT‑4V‑class, Gemini) where they're strong — interpreting tables, notes, and ambiguous annotations, and bootstrapping training labels — while keeping them out of the precision‑critical measurement path. What We're Looking For Strong Python engineering fundamentals; comfort owning a production pipeline end to end Solid computational geometry skills — line/polygon operations, spatial indexing, coordinate transforms, geometric pattern matching Hands‑on experience training, evaluating, and deploying detection and/or segmentation models, including dataset construction Experience extracting structure from PDFs, CAD files, or other complex document formats (e.g., PyMuPDF, pdfplumber, ezdxf, or comparable tooling) Experience with OCR and layout analysis pipelines (PaddleOCR, Google Document AI, or similar), or the judgment to integrate them well Pragmatism about ML: deterministic solutions when the data supports them, learned models when it doesn't, and clear reasoning about which is which Early‑stage temperament: bias toward shipping, comfort with ambiguity Nice to Have Background in drawing/map vectorization, document‑to‑structure problems, or graph‑based scene reconstruction Familiarity with LayoutLM‑style document AI models or Document AI platforms Experience building human‑in‑the‑loop labeling and continuous model‑improvement systems Familiarity with CAD internals or engineering document conventions Why This Role You'll own the company's central technical problem, working directly with the founder. It's a rare blend — computational geometry, document AI, applied deep learning, and relational reasoning in one pipeline — and everything you ship reaches real users immediately, feeding the loop that makes the models better. #J-18808-Ljbffr Girder AI