Flexiple
Website:
flexiple.com
Job details:
Hiring for COSO Intelligence — via Flexiple, the recruiting partner on this role.
About Company
- COSO Intelligence is a construction intelligence platform that uses AI to help construction professionals work with project data and make decisions.
- Company website: https://cosobuild.ai
Job Roles & Responsibilities
- Build production Python APIs and asynchronous workflows for reliable document-grounded AI services.
- Develop hybrid RAG retrieval, reranking, query rewriting, chunking and metadata filtering over production vector stores.
- Establish labelled AI evaluation datasets and CI/CD release checks for retrieval quality, groundedness and faithfulness.
- Enforce strict multi-tenant RLS and query-layer isolation across storage, retrieval, caching and background tasks.
- Deliver source-document/page citations and honest refusals when available evidence cannot support an answer.
- Build layout-aware extraction and OCR pipelines for messy documents while preserving provenance and page metadata.
Ideal Candidate Profile
- 3–5 years of experience with hands-on production backend and AI delivery.
- Own production Python services using FastAPI or Django, REST APIs and async execution; PostgreSQL, Redis, Celery and Docker. Show deployed workloads, queue retries/idempotency, performance and failure handling rather than notebook-only work.
- Ship hybrid dense plus sparse/BM25 retrieval with reranking, query rewriting, chunking and metadata filtering. Demonstrate a production vector store such as Qdrant, pgvector, Pinecone or FAISS; these are alternatives, not cumulative requirements.
- Maintain labelled evaluation sets in CI/CD; use Recall@k, MRR and nDCG for retrieval and groundedness/faithfulness for answers. Explain baselines, regression thresholds, dataset versioning and failure slices.
- Enforce tenant isolation through RLS at the database/query layer, not prompts alone. Carry tenant context through retrieval, cache keys and Celery jobs; demonstrate negative cross-tenant tests and least-privilege service access.
- Return source-document and page citations tied to evidence actually retrieved; refuse or qualify unsupported answers honestly. Explain citation provenance, unsupported-query tests and failure handling.
- Extract messy PDFs, scanned documents and complex layouts using layout-aware extraction/OCR; example tools include Tesseract, docTR, Textract and DocLayout. Preserve page identity, reading order, tables and metadata into retrieval.
- Preferred, not mandatory: Agent orchestration with human-in-the-loop using LangGraph, Agno, AutoGen or CrewAI; staged actions committed by a human and no unattended database writes.
- Preferred, not mandatory: Quantized self-hosted model routing using vLLM or Ollama, int4/int8, data sovereignty, open-source fallbacks and cost/latency tradeoffs.
- Preferred, not mandatory: OpenTelemetry (OTel), structured logging, tracing and operational dashboards.
- Preferred, not mandatory: PySpark, Parquet and CDC-based ETL.
- Preferred, not mandatory: AWS or GCP cloud delivery; LoRA/QLoRA only with judgement about when fine-tuning is unnecessary.
What We Offer
- Remote work.
- Hands-on ownership of backend and document-grounded AI engineering for construction intelligence.
Originally posted on: 12 Sept 2026
Click on Apply to know more.