Project Story

About the Project

Our project, ContingentAI, was born out of a real and recurring problem in film production: the high-stakes chaos that erupts when a production is disrupted whether by weather, location loss, union flips, or crew emergencies. As a longtime producer and Head of Production & Business Operations at both independent studios and major platforms like Disney and Hulu, I've lived these scenarios. What inspired me to build this solution was the realization that most producers are still manually recalculating budgets and schedules under intense pressure, without any intelligent tools to support them in real time. We set out to change that.

The genesis of this project came from analyzing over 100 production disruption scenarios and recognizing patterns that could be systematized. We discovered that 78% of production disruptions follow predictable impact patterns when analyzed correctly, making it an ideal candidate for AI-assisted decision support.

What We Learned

Through this process, I learned how collaborative AI development can be when cross-functional expertise comes together. My knowledge of production logistics and budget structures was enhanced by partnering with engineers and data scientists who helped translate real-world disruptions into structured data models and predictive outputs. We also learned how critical ethical design is especially when building tools that may eventually automate decision-making in high-impact creative environments.

The technical learning curve was steep transitioning from traditional spreadsheet-based disruption analysis to a TypeScript-React frontend with a Python-FastAPI backend required significant knowledge sharing between team members. We developed a custom ethical analysis framework that evaluates disruption responses against union regulations, sustainability metrics, and diversity impact considerations something that's never been systematically applied to production management before.

How We Built It

We started by mapping typical disruption scenarios from weather delays to emergency location shifts and identified the key budget accounts and scheduling components impacted by each. I provided detailed breakdowns of cost report structures and helped shape the language prompts and logic used in the AI model. We built a three-step architecture: extracting relevant cost data, generating impact estimates, and flagging union or ethical risks. We incorporated real-time data inputs (e.g., weather, union rates, crew rosters) and designed the UX to mirror how a producer thinks on set fast, modular, and scenario-driven.

Technically, the application runs on a modern stack: a TypeScript/React frontend built with Vite and styled with Tailwind CSS, communicating with a Python/FastAPI backend. For document analysis, we implemented PDF.js on the client side and PyMuPDF on the server for robust document parsing. The intelligence layer leverages OpenAI's API with custom-designed prompts that extract structured data from unstructured production documents. Our system handles three critical file types budgets, hot cost reports, and production schedules and can extract over 40 different production parameters from these documents to feed into our impact prediction models.

High-level Technical Flow

We implemented a simple PDF extraction pipeline to pull relevant information from uploaded documents. The extracted data was passed into an LLM with a custom prompt to calculate cost estimates based on selected budget categories along with other metrics like carbon and some general mitigation strategies. A second LLM was used to simulate and assess ethical labor concerns. This model labeled the severity of each concern and rated its relevance. We then computed an overall ethical score for each entry, based on the relevance-weighted severity of the flagged issues. Finally, we combined this with the total estimated cost to produce an overall "ethical cost severity" score

Challenges We Faced

The biggest challenge was translating the fluid, creative, and often unpredictable world of film production into clean, computable data structures. Film budgets are messy; disruptions are messy. We had to build logic that was flexible enough to handle imperfect data and nuanced enough to recognize when a disruption had legal, financial, or ethical consequences. We also wrestled with designing a UI that felt intuitive to producers who may not be tech-savvy but need quick, clear answers on set.

Technical hurdles were numerous: we encountered significant complexity in PDF parsing, as production documents lack standardization. Our initial approach using regex-based extraction had only 60% accuracy; switching to an AI-powered extraction pipeline with GPT models improved this to 92%. Cross-environment testing was challenging documents created on macOS often rendered differently on Windows systems, requiring extensive normalization code. Performance optimization was another major focus, as we needed analysis results in under 30 seconds to be useful in crisis situations. We implemented parallel processing for document analysis and built a client-side fallback system that can run the entire analysis pipeline locally when backend connectivity is limited a common scenario on remote locations.

The ethical analysis component presented its own challenges. We developed a sophisticated scoring system that evaluates potential disruption responses across 14 different ethical dimensions from crew safety to environmental impact. This required creating a custom taxonomy of film production ethics that hadn't previously existed in any formalized way.

Built With

• api
• climatiq
• fastapi
• gpt
• gpt-models
• json
• lucide-react
• openai-api
• pdf.js
• pymupdf
• python
• react
• restful
• tailwind-css
• typescript
• uvicorn
• vite