Applied AI Engineer (Document Intelligence & Prompting)

About

Hooh is an AI-first startup turning cutting-edge models into real products. We love turning research into usable tools that solve real problems. If you thrive on ownership and impact — you’ll fit right in.

We’re looking for an Applied AI Engineer to make messy documents reliably understandable — combining accurate structure and entity extraction with strong retrieval and cost-efficient model orchestration.

Key responsibilities

• Document analysis & taxonomy — analyse layouts, sections, entities, and relations; define categories, types, and metadata schemas that generalise across domains
• Prompting & schemas — author, version, and maintain prompts/chains for classification, extraction, summarisation, and Q&A; enforce JSON/JSON-Schema outputs
• Retrieval (RAG) — build hybrid retrieval (embeddings + keyword + graph traversal); manage embedding generation, indexing, deduplication, and drift
• Cascades & MoE — architect model cascades and MoE routing to balance quality, cost, and latency
• Agentic orchestration — coordinate multi-step pipelines with caching, batching, streaming, and robust retries
• Result fusion — combine outputs via reranking, voting, and confidence/consistency checks; implement guardrails and safety rules
• Evaluation & monitoring — run evals (LangSmith or similar), define golden sets, automate regressions, track quality/cost/latency; canary and observe
• Operate models locally or hosted — vLLM / TGI / Ollama / llama.cpp and API models; apply quantisation/LoRA when useful
• Data stores — vector DBs plus graph and relational DBs (Qdrant, Weaviate, Milvus, pgvector, Neo4j, PostgreSQL)

Requirements

• Strong programming in Python or JavaScript/TypeScript; clear, tested, maintainable code
• One LLM-powered feature shipped end-to-end (prototype → production), or equivalent open-source/portfolio work
• Hands-on with RAG, vector embeddings, and evaluation (offline + online A/B, error analysis)
• Familiarity with model cascading / MoE concepts (or ability to learn quickly)
• Practical database skills; comfortable with SQL and with graph or vector systems
• Product mindset with a bias to measure impact and iterate