doAZ

Senior AI Solutions Architect — Enterprise Knowledge Systems Doaz | Remote (Global, KST ±5h preferred) | 30~40 LPA About Doaz Doaz builds domain-expert vertical AI for construction, heavy industry, and public safety. Our products— ConGPT (construction knowledge RAG), DGPT (multimodal quality-manual chatbot), PointChecker (contract/spec analysis), and GeoAI-Suite (borehole & excavation analysis)—help organizations like Doosan Enerbility, POSCO E&C, KT Estate, Lotte E&C , and Seoul Fire & Disaster HQ reduce risk, compress review cycles, and keep people safe. The Opportunity Architect industrial-grade AI that turns PDFs, CAD/vector drawings, incident reports, and regulations into auditable, evidence-linked decisions . You’ll own end-to-end systems (SaaS & on-prem) that prevent accidents, ensure compliance, and save lives —at Fortune-500 scale. What You’ll Do 1) Enterprise RAG & Knowledge Architecture Design multimodal RAG over PDFs (text/tables/images), CAD/vector drawings, MSDS/chemicals, and bilingual (KOR/ENG) regulations. Implement hybrid retrieval (BM25 + dense + metadata + knowledge graph) with reranking; target ≥90% top-k answerability and traceable citations . Build domain embeddings for bilingual technical terminology; craft routing/prompting for audit-ready outputs. 2) Industrial Data & Change-Aware Pipelines Ingest/normalize heterogeneous sources (ERP/MES exports, legacy DBs, SharePoint, IoT streams). Ship 10k+ document-type ingestion with schema validation, redaction, and temporal versioning for regulatory drift. Integrate external APIs (e.g., KOSHA, OSHA, EPA ; for finance modules SEC/DART ). 3) Production AI & MLOps Orchestrate ensemble decisioning (rules + priors + ML) with SLA < 10s ; cost-optimized LLM flows (tool-use, caching, distillation). Operate on AWS EKS , Postgres, Pinecone, Temporal/Argo , Prometheus/Grafana; CI/CD with test & eval gates. Build explanation layers (attribution, chain-of-evidence; SHAP/LIME where applicable) and human-in-the-loop feedback. 4) Vision & Document AI (Nice to have) Table/figure/annotation extraction (LayoutLMv3/Donut/DocFormer), symbol detection ( YOLOv8/RT-DETR ), PDF vector parsing. 5) Client Co-Creation Lead deep-dive workshops with CxO/stakeholders; design PoCs that land $1M+ programs. Mentor client teams and internal engineers; author crisp technical docs fit for audits. What You Bring Must-Have 7+ years building production AI/ML or search systems; 3+ enterprise deployments end-to-end. Expert Python 3.11+ , SQL; strong systems thinking and data modeling. RAG at scale: vector DBs ( Pinecone/Weaviate/FAISS ), BM25, rerankers, prompt/routing strategies, evals (faithfulness, coverage, latency). MLOps/SRE: versioning, canaries/A-B, drift detection, observability, cost/perf trade-offs. Clear communication; ability to turn messy, multilingual data into reliable software. Nice-to-Have Knowledge graphs ( Neo4j/RDF/SPARQL ), schema alignment/ontologies. VL/Document AI (LayoutLMv3, Donut), CAD/vector parsing, safety/compliance domain exposure. Orchestration frameworks (Temporal/Argo), FastAPI , AWS EKS , PostgreSQL , Pinecone . LLM fine-tuning/LoRA, retrieval-graded generation, multi-agent planning. Our Stack (you don’t need all of it) GPT-4o/GPT-4.1, Gemma-3 27B-VL , Qwen-VL , rerankers (Cohere/MXBAI), YOLOv8/RT-DETR, LayoutLMv3/DocFormer/Donut, FastAPI, Python, TypeScript, AWS EKS , Pinecone, Postgres, Temporal, Prometheus/Grafana. Interview Process passed the (initial) resume screening Technical Screen (45m): Your RAG decisions; live triage of a retrieval accuracy issue. System Design (2h): End-to-end design for a compliance-grade knowledge system (data → RAG → UI/UX). Take-Home (48h window): See “One Question to Answer” below. Founder Conversation: Vision/values alignment; references with prior enterprise clients. One Question to Answer (Take-Home Challenge), Challenge: Multi-language Safety Document RAG Prototype Goal: Build a working RAG system that processes safety documents in Korean and English , returning evidence-backed answers (required: inline citations and source snippets). Dataset Provided: 50 safety incident reports (25 Korean, 25 English) 10 regulatory PDFs (5 Korean KOSHA , 5 English OSHA ) 20 MSDS sheets (mixed languages) 10 sample queries with expected answers Deliverables (48h window): A running API (or CLI) that answers the 10 queries with citations Brief README covering: indexing strategy, retrieval pipeline (hybrid choices), chunking, reranking, bilingual handling, and evaluation method Report with metrics : answerability, faithfulness (citation match), and latency (p50/p95) What we’re looking for: Sound architecture, multilingual retrieval quality, clean evidence chains, pragmatic cost/latency trade-offs, and a clear eval plan. How to Apply Email: doaz@doaz.ai Subject: [Senior AI Architect – YOUR_NAME] Include: GitHub or repo for a production RAG/search system you built 1-page architecture diagram of your most complex deployed AI system Concrete metrics (accuracy, latency, scale, cost) (Bonus) Live demo URL, industrial AI write-ups, OSS contributions Why This Matters: The systems you build will directly influence real-world decisions at scale—improving compliance, reducing risk, and protecting workers every day.