Senior CV / LLM Engineer

Job description

About Doaz

Doaz is a hyper-growth startup on a mission to turn fragmented industrial knowledge into instant, actionable insight. We build LLM- and Vision-AI solutions for construction, heavy industry, and finance leaders who must transform terabytes of drawings, specifications, and regulations into real-time decisions.

We’re expanding our GeoAI programs (including joint work with POSCO E&C) and launching drawing-change detection services that automatically compare plan versions, detect deltas, and explain design impacts.

Why You’ll Love Working Here

• True 0-to-1 ownership — Ship models that land in production sites within weeks.
• Global impact, lean crew — 30 teammates across KR / PK / IN; no bureaucracy, only builders.
• Tech freedom — YOLO or RT-DETR? Gemma-VL, Qwen-VL, or LLaVA? You choose, we fund.

Role Overview

We’re hiring a Senior Computer Vision & Multimodal LLM Engineer (GeoAI & Drawing Change Analysis).

You’ll lead end-to-end development of a version-aware drawing-diff engine (PDF / DWG raster & vector), symbol / text extraction, and change-impact narratives powered by RAG / LLM. Expect fast cycles from prototype → service : detection models, OCR / layout understanding, retrieval, and explainable outputs that engineers can trust.

Key Responsibilities

Drawing Change Analysis (CV)

Build a robust diff pipeline for architectural / structural / MEP drawings : rasterization, layer parsing, vector geometry ops, and semantic change clustering.

Train / finetune detectors & segmenters (e.g., YOLOv8 / RT-DETR / Detectron2 / SAM) for symbols (columns, openings, sleeves), title blocks, and revision clouds; achieve production-grade mAP / F1.

Implement geometry-aware post-processing (IoU / topology checks, snapping, graph connectivity) to reduce false positives.

Document & Layout Understanding

Engineer OCR + layout models (PaddleOCR / Tesseract + DocFormer / LayoutLMv3 / Donut) to read legends, notes, schedules, and BOM tables; normalize to structured JSON.

Build version-aware entity tracking (IDs, gridlines, BH IDs, coordinates) across revisions.

GeoAI & LLM / RAG

Design retrieval over drawings / specs (BM25 + vector) with reranking; ground LLM answers in evidence with citations and clickable locations.

Generate change-impact summaries (e.g., slab shear reinforcement, opening proximity to columns) with rules + LLM verification; measure factual precision.

Productization & DevOps

Ship FastAPI / gRPC microservices, batch & streaming workers (Ray / Celery), GPU inference (Triton / TensorRT), and observability (Prometheus / Grafana).

Own evaluation : dataset curation, data labeling guidelines, ablation / A-B tests, and regression suites.

Collaboration

Work closely with domain SMEs (geotech / structural) to encode rules (KDS / KBC, internal standards) and prioritize what matters to the field.

Minimum Qualifications

5+ years of production Python (3.x) building ML-heavy backends; strong PyTorch.

3+ years in computer vision for detection / segmentation / OCR or document AI at scale.

Hands-on with multimodal LLM / RAG (LangChain / LlamaIndex), vector DBs (Pinecone / Weaviate / FAISS), and rerankers.

Proven experience parsing engineering drawings or complex PDFs (vector / raster), including geometry and layout reasoning.

Solid MLOps : reproducible training, CI / CD, model packaging, monitoring; cloud on AWS / GCP.

Fluent written & spoken English (Korean a plus).

Preferred Extras

GPU orchestration (Kubernetes / Ray / Slurm), high-performance inference (ONNX / TensorRT).

Experience with VLMs (Gemma-VL, Qwen-VL, LLaVA), CLIP, or doc-layout models.

Open-source contributions, papers, or strong public demos in CV / doc AI / RAG.

Full-stack chops (TypeScript / Next.js / React) for quick operator tools and review UIs.

Compensation & Benefits

Competitive base salary (market-leading) , around 20 lakh (yearly)

Performance-based annual bonus (up to 20%).

cloud credits, and AI tools support.

Hiring Process (≈ 2–3 weeks)

Quick intro call (15 min, mutual fit).

48-hour take-home : Drawing Diff + Evidence-Grounded Summary (provide code + short README; clarity >

polish).

Deep-dive tech interview : architecture, modeling choices, evaluation, and scaling plan.

Culture & vision chat with Founder / CEO.

Offer — if all green, written offer within 24 h.

How to Apply

Email doaz@doaz.ai with subject [CV / LLM Engineer – Your Name] and include :

Résumé / CV with measurable outcomes (metrics, latency, cost, accuracy).

Current or recent salary.

GitHub and / or live demos of CV / doc-AI / RAG work (links preferred).

A one-page diagram of your “Drawing Revision → Detection → Evidence → LLM Narrative” pipeline, noting models, retrieval, and evaluation metrics.

Employment type : Full-time

Ready to turn messy drawings and specs into instant, trusted intelligence?

Let’s build the future together at Doaz.