Python developer
About the project & role
#497 Python developer English: C1 | Advanced -, C2 | Advanced +, Native Duration: 12 Dedication: Both Vacancies: 4 Project description The project is to design, build, and operate production-grade agentic AI systems for NBCUniversal that automate meaningful digital and broadcast media workflows. The work covers an agentic harness (tool use, multi-step reasoning and chaining, streaming, skills, multimodal integration, RAG, sandboxing, and state management), backend APIs and services, and large-scale batch workflows powered by foundation models (LLMs, VLMs, and open-source models). The team is also responsible for search and retrieval over rich media content, for evaluation frameworks that measure quality, reliability, and safety at scale, and for the user-facing experiences that expose these capabilities. The system must handle non-deterministic model behavior through observability, human-in-the-loop review, and graceful failure handling, and must balance quality, latency, cost, and explainability. The expectation is production delivery, not prototypes, in an Agile environment. Main tasks The candidate will architect and build the agentic harness from scratch and integrate multiple agent frameworks where appropriate. They will design and operate RESTful, streaming, and asynchronous APIs and services, choosing between synchronous calls, queues, event-driven patterns, and WebSockets based on latency, reliability, and scale. They will build and orchestrate large-scale batch pipelines using LLMs, VLMs, and deep learning models over video and rich media (using tools such as FFmpeg and OpenCV). They will design and tune lexical and embedding-based search (OpenSearch/Elasticsearch), covering relevance, ranking, indexing, and retrieval performance. They will create and maintain evaluation datasets and run offline regression and online production evaluations using rubric scoring, AI-as-judge, and human review. They will also own production reliability (monitoring, uptime, defect resolution, and roadmap contributions) and build front-end experiences in React/TypeScript. Finally, they will produce architecture artifacts (C4 model and related designs), collaborate with the client’s broadcast and media stakeholders, and make documented trade-off decisions on model selection, prompting, and sampling parameters. Required skills Agentic systems: 6 to 12 months building tool-using, conversational, long-running, or planning agents, including a custom harness with state, RAG, sandboxing, and streaming. Fluency with frameworks such as OpenAI Agents SDK, LangGraph, Google ADK, or SmolAgents. Backend engineering: 7+ years in Java, Go, Python, or JavaScript, and 5+ years shipping production software, including APIs, async workflows, queues, versioning, authN/authZ, and observability. AI evaluation: 1+ year evaluating AI systems for quality, reliability, and safety, with eval datasets, regression testing, AI-as-judge, and precision/recall and statistical significance analysis. Search and retrieval: 2+ years architecting search at scale, with lexical and embedding-based methods, ideally OpenSearch or Elasticsearch. Foundation model knowledge: use of LLMs and open-source models beyond API calls, including temperature, top-p, logprobs, prompting strategies, and model selection trade-offs. Architecture and process: C4 model and traditional design artifacts, Agile experience, and a BS in CS, EE, or equivalent experience. Desired skills Multimodal and media: experience with video, rich visual media, or multimodal AI, plus expert knowledge of digital and broadcast production workflows. Training or fine-tuning models for domain-specific use cases (LoRA, PEFT). Stateful design patterns and streaming protocols such as WebSockets. Front-end technologies (React, Angular, JavaScript, CSS, HTML). 5+ years of AWS and 1+ year of Azure experience. CI/CD practices. Docker and Kubernetes. Media-processing libraries such as FFmpeg and OpenCV.