ABOUT LUMEN
A better way to decide
what gets made.
Lumen turns fragmented creative material and live market evidence into a calibrated, explainable outlook for film and television teams.
THE PROBLEM
Greenlight decisions are expensive.
The evidence is fragmented.
Studios weigh scripts, talent, audience behavior, market timing, comparable titles and creative instinct—often across disconnected teams and tools.
Lumen does the repetitive research and calibration work in parallel, while keeping the final reasoning visible to the people making the call.
SYSTEM ARCHITECTURE
Agentic training.
Agent-directed recommendations.
Two connected workflows: an agentic pipeline trains the ML model, then a main agent uses that trained model as a deterministic tool alongside comparison and research tools.
01 · Agentic ML training
Trained once today · automation plannedThe training agent guides feature engineering, model selection and evaluation through a pipeline. Today we run this process once; future automation will trigger and manage repeat training runs.
02 · Runtime recommendations
Agentic orchestration · deterministic ML inferenceThe main agent receives screenshots, frames, story text and project context through the prediction API. It extracts structured features and decides which tools can run concurrently. Calls with dependencies wait for the inputs they need.
WHY AGENTS
The agent chooses.
The model predicts.
Training decisions and runtime tool selection are agentic. The trained ML model is a deterministic tool: the same validated features and model version produce the same prediction. The agent’s plan and final explanation can vary with the request and available evidence.
Read screenshots, frames and text into structured features that match the trained model’s input schema.
Select the model, comparison and research calls needed for this request; run independent calls concurrently.
Join tool results and assess similarities, differences and conflicting evidence.
Explain the recommendation using model predictions, comparison findings and relevant market evidence.
BUILT FOR AGENTIC CINEMA
Each component has a clear role.
The main agent extracts features from creative materials and directs runtime tool use.
A pipeline trains and evaluates the ML model. It is run once today, with automated repeat runs planned.
Parallel Search is a research tool. Concurrent execution is a separate scheduling decision made by the main agent.
The prediction API exposes the trained model as a tool; inference uses a fixed model version without retraining.
THE TEAM
Built in Stockholm.
Made for the screen.
Three builders working across product, cinema, data and engineering.
