Projects using Airtable in Little RockProjects using Airtable in Little RockAdapting Dungeon Crawler Carl into a Multi-Episode AI Video Series
I am currently producing a full-length, multi-episode video adaptation of the bestselling LitRPG series Dungeon Crawler Carl by Matt Dinniman, built entirely using AI video generation tools.
This is not a proof of concept or a short demo reel. It is a long-form narrative production where character consistency across every scene, every episode, and every camera angle is the central technical challenge.
The release is still roughly two years out. What I can share now is the production methodology, the pipeline architecture, and a sample of the visual quality the system produces.
https://media.contra.com/video/upload/fl_progressive/w_1600/h0sumzd0faydmt1eksbj.mp4
Sample output from the Kling AI 3.0 pipeline. Audio omitted due to proprietary content.
The Core Problem: Character Consistency at Scale
Single-shot AI video is easy. Maintaining a character's face, build, wardrobe, and physical presence across dozens of scenes, camera angles, and lighting conditions over multiple episodes is where most AI video projects fall apart.
I solved this using a combination of strict reference management, programmatic prompting, and a locked-string discipline that treats every character description as an immutable asset.
Production Architecture
The pipeline operates as an interconnected automation ecosystem, not a single tool.
Pre-Production — Midjourney / Flux
Every character begins as a locked "Golden Sample": a master reference sheet with four clean angles (front, back, side profile, and 3/4 view). These feed directly into Kling's Element Binding system so the AI never has to guess an unseen angle. This is the single most important step for preventing "face melting" across shots.
Rendering — Kling AI 3.0
Kling is the core rendering engine. Version 3.0 supports multi-shot generation (up to 15 seconds with 6 defined camera cuts per block), native audio synchronization, and Subject Element Binding for character consistency. I operate it like a physical soundstage, not a chatbot.
Audio — ElevenLabs
Highly customized emotional voice cloning and clean dialogue tracks. Kling's native lip-sync maps directly to external audio files.
Orchestration — Make.com (http://Make.com) + Airtable
Because rendering 1080p clips takes several minutes per shot, the pipeline uses asynchronous webhooks to manage API handoffs between Airtable databases and Kling's servers. Airtable acts as the master database for shot lists, character lore, and automated tracking matrices.
Post-Production — CapCut + Topaz AI
Micro-batched 4-to-5-second clips are stitched together in CapCut, with timeline pacing, universal color LUTs, and final 4K upscaling through Topaz AI. Raw assets (15MB to 30MB per clip) are mirrored locally via desktop sync to prevent cloud-fetching latency during editing.
Prompting as Directing
The quality gap between amateur AI video and production-grade output comes down to how you write the prompt. I treat every prompt as a visual environmental tracking matrix.
Time-Coded Storyboarding
Kling responds best to sequential, time-coded directives. Transitions are forced using hard time markers (e.g., "Shot 1 (0-4s): Wide establishing shot...") rather than narrative paragraphs.
Motivated Camera Moves
Every camera movement gets a strict physical path: a slow dolly push, a rack focus, a locked tracking shot. Unmotivated camera movement is the primary cause of character mutation in AI video.
The Locked String Discipline
When describing a character's state or gear, the exact same string of words is used in every single generation. Changing "rugged combat gear" to "dirty armor" forces the AI to recalculate the asset from scratch, causing visual drift. Consistency is a vocabulary problem as much as a technical one.
Kinetic Anchors
To break the smooth, weightless "AI look," I introduce natural forces into every scene: wind, dust, friction, fabric weight. This forces the engine to calculate environmental resistance and real-world physics, producing motion that feels grounded.
Why This Matters
AI video production is moving fast, but most of what exists today is short-form, single-character, single-scene content. Building a multi-episode narrative with consistent characters, coherent world-building, and production-grade visual quality requires the same discipline as traditional film production, just with a fundamentally different toolchain.
The methodology I have built for this project applies directly to any long-form AI video production: brand series, product narratives, educational content, or entertainment.
This project is in active production. More will be shared as the release approaches.
Check it out (https://contra.com/p/kmmRerLW-ai-video-production-bringing-a-bestselling-book-to-life) Building an AI-powered affiliate publishing engine from the ground up.
I am currently working on a full automation system for NodeRidge, designed to turn product and software research into structured, review-ready WordPress content with human approval built into the workflow.
The goal is simple: remove the repetitive work from affiliate publishing while keeping quality control, visibility, and editorial review in place.
This build connects:
🗄️Airtable as the central operations hub
⚡Make.com (http://Make.com) for workflow automation
🧠OpenAI for content generation
📝WordPress for draft publishing
📊Google Search Console for rank and performance tracking
What I like most about this project is that it is not just "AI writing content." It is a real backend publishing engine with structured data, review gates, sync monitoring, and operational visibility.
The system is already handling prompt runs, WordPress draft syncing, affiliate link registry logic, and QA tracking. My next focus is scaling the Airtable architecture to handle higher volume.
This is the kind of build I enjoy most: connecting messy manual workflows into a clean system that can actually scale.
What are the biggest operational bottlenecks you're currently dealing with in your own workflows? Stop building science experiments and start building scalable, autonomous operational systems.
I just published a detailed look under the hood at my most sophisticated AI integration yet: The Hierarchical Multi-Agent Business System.
Most founders use ChatGPT to write emails. I build AI agent teams to source leads, design technical architectures, manage client-facing dashboards, and automate complex content pipelines—all with zero human input for everything except final approval.
Take a look at the attached infographic to see how this recursive, self-optimizing architecture functions. Here is the operational impact of what I build:
Departmental Specialization:
I don't use one "smart" agent. I build a crew of highly specialized agents (Lead Gen, Content, DevOps) overseen by a Manager Agent to ensure quality and prevent hallucinations.
The Human-in-the-Loop Safeguard:
The entire workflow is designed to be fully autonomous until final quality control and strategic approval. You get the scale of AI with the certainty of a human final check.
Tool Agnosticism:
My builds are logic-first. We use the best tool for the task, whether that's CrewAI, LangGraph, Supabase, Cloudflare, OpenAI, or a custom API.
If you are a founder or operator ready to turn your standard business workflows into high-performance, autonomous engines, check out the full case study.
https://contra.com/s/2rFLgNAU-custom-ai-agent-for-your-business-workflow 
