EXPERIMENT_005 // AUTONOMOUS.BRAND.PIPELINE
Submit a creative brief and brand rules. A three-stage LLM pipeline generates N concept variants, evaluates each against your rules, and ranks the top K — streamed live as it runs.
LOADING EXPERIMENT...
A three-stage agentic pipeline runs entirely on the server: a generate step produces N brand concept variants from a brief and brand rules, an evaluate step scores each variant for rule compliance, and a rank step selects the top K results. Every LLM call streams token-by-token over SSE so you watch the pipeline think in real time — not a progress bar, but the actual model output.
HOW IT WORKS
Submitting the form POSTs a config to the API, which enqueues a BullMQ job backed by Redis and returns a jobId immediately. A long-running Node.js worker picks up the job and runs the three pipeline steps sequentially. BullMQ handles retries, job isolation, and worker lifecycle — the HTTP response returns before a single LLM token is generated.
The frontend connects to a streaming endpoint using the jobId. The server publishes structured events — step_start, llm_chunk (one per token), step_complete, pipeline_done, pipeline_error — to a Redis pub/sub channel. The SSE handler subscribes and forwards each event to the client as a typed JSON frame. The stream closes automatically on terminal events.
Generate: the model receives the brief and brand rules and returns N distinct brand concept variants as a JSON array. Evaluate: each variant is scored independently against the brand rules on a 1–10 scale, with flags for specific violations and a rationale. Rank: all scored variants are compared holistically, ties broken by flag count, and the top K are returned in rank order.
A built-in Luxury Sustainable Fashion preset demonstrates the rules format — voice, colour palette, typography direction, values, audience, and explicit avoid-list. The preset is a one-click fill for the brand rules field, designed to be edited rather than used verbatim. Writing precise rules produces tighter evaluations: vague rules produce vague scores.
WHAT THIS PROVES
The autonomous execution model decouples configuration from observation. The user provides constraints once, then watches a multi-step agent pipeline execute server-side without further interaction. Each step — generate, evaluate, rank — involves Claude reasoning on structured data, not just prompt-response pairs. The pipeline demonstrates that real agentic work in a web application context is not polling for an answer; it is delegating work to a persistent job queue and observing the reasoning unfold in real time. This is the pattern for production agent systems: intent specification, autonomous execution, full observability.
The stack itself is the infrastructure lesson: BullMQ provides job persistence and concurrency control; Redis pub/sub provides a low-latency observation channel; Anthropic SDK's streaming callbacks provide token-level visibility into Claude's output; SSE delivers those tokens to the client without polling. All four layers are required for production agentic work at web scale. A job queue without streaming gives you answer-on-demand but no visibility. Streaming without a job queue couples the client to the inference duration. This experiment shows why that coupling breaks down and how to separate them.