Animation study
Animation-only render
An early visual-system pass focused on the motion and editorial layer. The narration is my own voice, generated through ElevenLabs.
Agentic Systems · Case Study
TC_ACC: an autonomous documentary studio
One case reference link goes in. Roughly five hours later, the system returns a finished ten-minute documentary package. This page presents the public scope and verified output without documenting the private prompts, tuning, routing, or production control logic.
28
Stages
5
Departments
2
Review gates
~5h
Current runtime
0
Human steps
What it produces
The public claim is deliberately simple: a case reference is turned into a researched, written, narrated, illustrated, edited, rendered, and packaged documentary episode. The four outputs below show the system at different points in its development.
Full pipeline
First full production pass
A longer end-to-end render from the same case, showing the assembled documentary package.
Full pipeline
Refined production pass
A later full-system iteration with the visual, narrative, audio, and finishing layers together.
Latest output · Oct 2026
Latest episode
A new case, researched, written, narrated and rendered end to end by the current pipeline.
Current runtime: a complete run takes about five hours. Time was not a project constraint, so I prioritized output quality and a reliable sequential workflow over concurrency optimization. That runtime is a current engineering tradeoff, not a speed claim.
A useful next test
Give the pipeline a case.
Have a story or unusual case you want to see processed? Leave your email and a short reference. The request is sent directly to me; a full pipeline run currently takes about five hours.
Public system overview
The current pipeline has five production departments, twenty-eight stages, and two independent review gates. The diagram stays at that published level; internal stage contracts, routing, retry policy, and tuned orchestration are intentionally omitted.
Research
Source intake, research, and evidence assembly
Story
Story development, script, narration, and voice
Assets
Visual planning, acquisition, and review
Independent review gate
Animation
Editorial assembly and motion
Independent review gate
Publishing
Branding, rendering, metadata, and delivery
Public evidence
Selected production codeThe excerpt contains real, low-disclosure-risk modules rather than display-only pseudocode.
46 published testsThe included modules and their tests run independently of the private package.
Measured outputThe published reference run reports 9:55 at 1920×1080, 14,294 frames, 128 shots, and 760 reviewed assets.
Public technology boundary
Python 3.12 for the pipeline, TypeScript for rendering · LangGraph and Pydantic for the public architecture · Remotion and ffmpeg for video · pytest for the published tests.
Named model and voice providers:
OpenAIGeminiQwenElevenLabsFish Audio
Specific models, routing order, provider contracts, fallback policy, and configuration are intentionally not documented here.
Disclosure boundary
The public repository is intentionally not a runnable copy of the production system. It includes selected modules, tests, redacted configuration structure, and explanatory docs; it omits the entry point, dependency manifest, environment template, and most production modules.
Intentionally withheld. Prompts, tuned configuration values, production orchestration, resume and reuse machinery, provider integrations, research/search logic, the brand and animation system, and most of the TypeScript renderer and agent implementation.
The portfolio follows the same boundary. It demonstrates the system's scope and public evidence without reproducing the transferable implementation details.