Beyond “Because You Watched”
Build a Movie Recommendation Agent with GraphRAG
How students use Knowledge Engineering to turn a general AI agent into a streaming solution
Imagine that you have just joined a leading AI consulting firm and have been assigned to a streaming provider as a Forward Deployed Engineer. Your first assignment sounds familiar: build an AI agent that helps people decide what to watch.
The client already has a large movie catalog, viewing histories, audience profiles, ratings, and content descriptions. It may also use recommendation models to produce suggestions such as “Because You Watched.” What it lacks is an agent that can hold a conversation, understand several viewers at once, explore relationships across the catalog, and explain why a recommendation fits.
Consider Jason and Anjali. They watch a movie every Friday—sometimes alone, sometimes with their twin nephews, and occasionally with friends. Jason enjoys the tension and competitive energy of sports stories. Anjali prefers dance movies. Tonight they want something they can enjoy together.
Their request is more specific than “show me a popular dance movie”:
“Can you find a movie that combines dance with the excitement of competition?”
One possible answer is Dirty Dancing: Havana Nights, depending on age suitability, availability, and the provider’s catalog. The point is not whether a language model knows the title. The challenge is to produce a grounded recommendation from the provider’s own content, audience knowledge, and a custom Knowledge Graph harvested from that content.
Movie Recommendation - User Persona
The project begins with the viewers, their situation, and the decision they are trying to make.
This is the central project in Build and Deploy Agentic AI Solutions: Agentic AI and Forward Deployed Engineering. Students take the use case from discovery through knowledge engineering, implementation, and continuous improvement.
Why a general-purpose agent is not enough
A general-purpose agent such as ChatGPT can already do two useful things. It can draw on public knowledge, and it can use a flipped interaction: asking questions before offering an answer. Who is watching? What mood are you in? Are there age restrictions? Do you want something familiar or surprising?
That conversational ability is a strong starting point. But the agent cannot automatically access the provider’s private catalog, regional licenses, viewing profiles, content policies, or current availability. It also does not know the classifications the provider has developed from its content and audience data.
A conventional recommendation model has the opposite strength. It uses platform behavior to suggest titles related to earlier viewing. “Because You Watched” works well when the next choice resembles the last one. It is less helpful when several people have different tastes, the occasion changes, or viewers want to understand why a title was selected.
When Jason and Anjali watch together, the system must connect two interests: dance and competitive sport. When their nephews visit, age suitability becomes a firm constraint and the choices shift toward children’s topics. When friends join, broad appeal and a shorter runtime may matter more. The right recommendation depends on audience context and relationships—not only on similarity to viewing history.
The missing middle is the Knowledge Layer
The course introduces Knowledge Engineering as the bridge between a reusable AI platform and the content harvested for each client. Across multiple engagements, a platform engineer can harvest a common Knowledge Graph that an FDE can customize for each customer.
For a streaming provider, the Knowledge Layer represents movie content and audience segments. Major content stores already use broad genres such as action and adventure, science fiction, drama and suspense, horror, and children’s movies. These categories are a useful starting point, but they cannot describe every viewing need.
Generative AI makes finer audience segments and movie categories practical. A Knowledge Graph can link actors, keywords, themes, and viewing contexts. Those links are flexible: Jason likes competitive sports, but would he also enjoy the competition in a ballroom dance movie?
Canonical entities such as movies, people, audiences, genres, themes, awards, and evidence sources
Relationships among movies, actors, and audiences
Custom categories such as Ballroom Dancing, Competitive Sports, and Family Viewing
Age suitability and movie ratings
The platform team owns this Knowledge Graph. It protects stable definitions, shared schemas, policies, and reusable components. The FDE works with the streaming provider to map its content into the architecture, identify gaps, and test extensions. The platform team can then review a useful new category or relationship and add it to the shared platform.
This shared architecture prevents two common failures: a product platform disconnected from real users and a separate, incompatible solution for every customer.
Start with the FDE use case
Students begin by mapping the audience and its needs. They identify the primary users, viewing context, business goals, available systems, and decisions the agent may make.
They then ask what would make the experience genuinely valuable. How should the agent engage viewers? What would delight them? What is the benefit to the provider?
These answers shape the design before anyone starts building.
Acquire the provider’s knowledge
The first Knowledge Engineering phase assembles the material the agent will need. To simulate a streaming provider, the course uses public datasets and MCP servers. The University of Minnesota’s MovieLens dataset and a TMDB MCP server act as surrogates for a provider’s content systems.
Students begin with MovieLens records, ratings, and user-supplied tags. They can enrich that data through an approved source such as TMDB, adding descriptions, runtime, release date, language, popularity, vote count, cast, and production information.
The provider can add private information that a public model cannot know: which titles are licensed, where they are available, household preferences, and the provider’s own content advisories. Current documents can also add information about recent awards or promotions.
Students learn to harvest a Knowledge Graph from this content. The graph may combine structured facts—such as runtime, release year, and popularity—with unstructured material such as plot descriptions.
Consolidate the knowledge around shared definitions
Raw source data rarely reflects how viewers talk about movies. A conversation about a historical film may focus on the events behind it. For a ballroom dance film, viewers may care about specific dance styles. Comedy calls for a discussion of humor, while action and adventure may be judged by their scenes and pace.
The Knowledge Graph exercises encourage students to create custom genres with precise descriptions, examples, borderline cases, and a recommended conversation. This gives the agent a consistent way to classify movies and guide the conversation instead of improvising every response.
Generative AI can uncover semantic similarities across movies. Similarity within the James Bond or Harry Potter series is easy to see. A harder question is whether The Pink Panther belongs closer to James Bond or The Three Stooges. Similar titles form a network, and connected groups can reveal useful categories missing from the original catalog. Course exercises include clusters such as “Heist and High-Stakes Crimes” and “Global Espionage Thrillers.” The same method can reveal a cluster around dance, performance, rivalry, and competition.
Semantic clustering helps the FDE discover candidate relationships and categories that the core team can review.
The resulting Knowledge Graph can connect Jason with competition themes, Anjali with dance, and candidate movies with both concepts. It can also connect each title to actors, awards, content attributes, and evidence passages.
Create the agentic consumption point
With the Knowledge Graph in place, students design the agent that will use it.
The agent combines conversational skills with tools that query and discover movies through GraphRAG. It starts by asking focused questions about the audience, occasion, preferences, and constraints. A preference skill separates firm requirements from softer interests.
GraphRAG then finds movies connected to both dance and competition. RAG retrieves supporting descriptions from approved sources. A ranking step considers relevance, audience fit, ratings, popularity, and variety. Before the agent responds, a validation step checks the evidence, freshness, and policy requirements.
The final answer should offer a short list instead of an endless feed. For each movie, the agent explains why it fits, identifies a tradeoff, and states its confidence. If the provider does not carry a title or the evidence is weak, the agent should say so and ask another question.
The agent uses Knowledge Graph descriptions to explain the chosen movie and each alternative through its dance styles.
Learn from the use
The first deployment will expose gaps. Viewers may use “sports” to mean competition rather than athletics. A dance taxonomy may need to distinguish performance stories from contest stories. A content advisory may conflict with a source description. These discoveries matter only if the team captures them systematically. The FDE monitors accepted recommendations, failed constraints, unsupported explanations, latency, and cost.
Together, those signals create a disciplined learning loop:
Customer interaction → field evidence → reviewed knowledge change → reusable platform improvement
The architecture remains stable enough for product teams to build on while still incorporating what field engineers learn from real viewers.
One project covers the complete lifecycle
The movie recommendation project follows the course’s FDE lifecycle:
Discover: Understand the streaming provider, its viewers, workflows, data, and measures of success.
Design: Establish the agent architecture, integration points, and controls.
Build: Prepare the movie data, taxonomy, graph, retrieval workflow, predictive signals, and agent.
Integrate: Connect the provider’s catalog, profiles, APIs, permissions, and user experience.
·Learn & Improve: Test with viewers and fit the agent into the provider’s discovery experience.
Build the solution in the October cohort
Build and Deploy Agentic AI Solutions: Agentic AI and Forward Deployed Engineering starts October 20. The four-week course includes four lectures, four office-hours support sessions, and hands-on work using both the zero-code Agentic Sandbox and code-first approaches.
Students learn the four-phase Knowledge Engineering framework: Acquisition, Consolidation, Agentic Consumption, and Learning and Governance. For the final project, each student acts as an FDE and takes a customer use case from discovery through architecture, prototype, evaluation, and learning. Successful participants earn an Intermediate-Level Agentic AI Certification.
You will choose your own use case and go through the agent building process similar to how I showed the movie recommendation example. The broader skill is learning to connect a reusable AI platform with customer-specific knowledge, build an agent in a real workflow using pre-defined Agent Architecture.
Comments