Senior Agentic AI Engineer

Company Description Imaginario AI is building a video intelligence layer that powers content discovery and storytelling for marketing and post-production teams in video-heavy industries. The company focuses on unlocking insights from video to help teams work faster, create more effective content, and streamline workflows. Imaginario AI collaborates with forward-looking creative and media organizations that rely on rich video libraries. The company has been recognized with the IABM BaM Awards for Content Management and Best of Show at IBC in both 2024 and 2025, highlighting its innovation and impact in the media technology space. What we're looking for Imaginario AI is looking for a Senior Agentic AI Engineer who has built and shipped LLM-powered systems in production and cares about making them accurate, reliable and affordable at scale. You specialize in LLM and agentic pipelines, and you are also a capable full-stack engineer who can take a feature from backend service to user interface. You will join a small, senior team and work closely with our engineering and product leadership to build the agentic layer of our video intelligence platform: agents that search, reason over and assemble video using our shot-level index, retrieval pipelines that ground every answer in real evidence, and MCP integrations that let AI assistants and third-party tools work directly with our customers' video libraries. This is a hands-on build role, not a research role. We care about what you have put in front of real users, how you measured it and what you changed when it went wrong. What you will do Agentic pipelines • Design, build and run multi-step agentic workflows in production: planning, tool and function calling, routing, and orchestration across several tools and data sources. • Coordinate several specialized agents that hand work back and forth, not only in a fixed sequence. • Make agents dependable with structured outputs, schema validation, retries, timeouts, fallbacks, guardrails and human-in-the-loop steps where the stakes call for them. • Build long-running agents with durable execution: checkpointed state, resume after failure, pause for user input, and branching to try alternative versions. • Keep agents fast and cost-efficient by tracking latency, token usage and cost per task, and by choosing the right model for each step. Context and memory management • Own context engineering: decide what goes into the model's context at each step, where it comes from and what gets cut when the budget is tight. • Compress and summarize long inputs such as hour-long transcripts and dense shot-level metadata without losing the facts that matter. • Build memory for agents and users: short-term session state, long-term memory across sessions and projects, and clear rules for what persists, what is retrieved and what is pruned. Retrieval-augmented generation (RAG) • Build and improve RAG over multimodal video metadata, including transcripts, shot descriptions, entities, topics and chapters. • Own the full retrieval stack: chunking strategy, embeddings, hybrid search (dense plus keyword), metadata filtering, query rewriting and reranking. • Diagnose retrieval failures and fix them at the root, whether the problem is in ingestion, indexing, ranking or generation. MCP and integrations • Build and maintain MCP servers that expose Imaginario's search, indexing and assembly capabilities to AI assistants and external agents, including tool design, schemas, async jobs with clear progress and status responses, authentication, rate limits and versioning. • Integrate third-party tools and data into our agents through MCP clients, REST APIs and webhooks, including media storage, MAM/DAM and editing systems. Agentic web search • Integrate third-party web search and agentic search APIs into our pipelines, for example to enrich asset-level metadata with trusted external information. • Handle source selection, freshness, deduplication and citation, and combine web results with internal retrieval so answers stay grounded and traceable. Accuracy, hallucination reduction and evaluation • Reduce hallucinations and improve accuracy with grounding and citation enforcement, verification and self-critique steps, constrained generation, and the ability to abstain when the evidence is weak. • Build our evaluation harness: golden datasets, retrieval metrics (recall, precision, MRR), faithfulness and answer quality checks, and LLM-as-judge scoring calibrated against human review. • Evaluate agent trajectories as well as outputs: tool choice, number of steps, cost and latency per task. • Run evals as regression tests on every release, and add tracing and observability in production so issues are caught by us before customers see them. • Turn user edits, ratings and accept or reject signals into better outputs over time, through rubrics, example-based context and eval sets. • Write clear documentation and share what you learn with the rest of the team. Full-stack delivery • Build features end to end, from Python backend services and APIs to the React interfaces where users work with agents, search results and assembled video. • Contribute to the reliability, performance and security of the wider platform, including services, background workers and cloud infrastructure. What you might work on in your first six months • Agentic workflows on top of our video index that turn a natural-language brief into searches, shot selections and an assembled first cut. • Our MCP connector, so customers can search and work with their video libraries from the AI assistants they already use. • An evaluation harness that tracks retrieval quality and hallucination rate across releases. • Agentic web research that enriches asset summaries with verified third-party title and entity information. Requirements • 5+ years of professional software engineering experience, including at least 3 years building LLM-powered applications that rea