Senior AI Application Engineer

Описание

You will build the applications that sit on top of our language models: agentic workflows, tool integrations and retrieval. You will also own the quality bar - from evaluation design to the release gates that decide whether a feature ships.

Отговорности

• Build agentic features end to end: tool contracts, routing, memory, streaming, retries and fallbacks, human approval gates
• Build and maintain tool integrations for internal systems, with per-user permissions and tight context discipline
• Design retrieval properly: ingestion, chunking, embeddings, hybrid search, reranking, metadata filters, freshness
• Own prompt and behavior control: versioned prompts with rollback, personas, output contracts, guardrails
• Own LLM evaluation: golden sets, automated judges, regression gates in CI, quality dashboards
• Ship the Python services behind it - asynchronous APIs, streaming responses, tested and documented
• Treat every tool response and retrieved document as untrusted input: prompt injection, taint, least privilege, sandboxing

Изисквания

• 5+ years building production Python services and APIs — not notebooks
• Hands-on LLM application work in production: tool calling, agent loops, retrieval and prompt systems, with users who complained
• Strong retrieval fundamentals, and the ability to measure retrieval quality
• Real evaluation practice: you have built a golden set or an automated judge and used it to block a release
• Enough transformer understanding to debug behaviour and cost — context limits, sampling, KV cache
• PostgreSQL plus a vector database (Qdrant, pgvector, Weaviate, Milvus) and Redis; Docker, tests, CI, code review

Preferred Qualifications

• Model Context Protocol, or deep knowledge of another tool protocol or agent runtime
• Reranker and hybrid-search tuning, with measured retrieval gains
• LoRA or QLoRA supervised fine-tuning with a before-and-after evaluation
• Speech, OCR or document pipelines (Whisper, PDF and DOCX extraction)
• Enough frontend to ship your own feature end to end (Next.js or React)

Technical Stack

• Application: Python, Flask, FastAPI.
• Retrieval: a vector database (Qdrant, pgvector), embeddings, rerankers
• Data: PostgreSQL, Redis, MongoDB, ClickHouse
• Models: internal LLM APIs, OpenAI-compatible
• Operations: Docker, systemd, nginx, Prometheus
• Frontend, when needed: Next.js, React