Marinara · Working animation pipelineTurning an ongoing roleplay into animated scenes.
I contribute to Marinara Engine and Marinara Agents, a local-first AI roleplay and game platform. My recent work connects persistent story context to a working MiniMax H3 / ComfyUI animation flow.
The basic idea is simple: roleplayers already create a script through dialogue and narration. The system can use that same material to produce a storyboard, visual references, keyframes, and animation without treating every shot like an unrelated generation.
How the pipeline works
- Use the roleplay as the script. Dialogue and narration are converted into a storyboard: a scene-by-scene animation plan.
- Ground the characters. Character-sheet artwork and persona data provide visual and written references so each shot starts from the same identity.
- Ground the locations. World Maps artwork provides a persistent visual reference for recurring places, including existing fictional settings.
- Create keyframes. AI prepares the still frame that anchors each shot, then refines the motion instructions around what actually rendered.
- Animate through MiniMax H3. ComfyUI combines the grounded frame and motion plan and sends them into the local H3 video workflow.
What works now: the H3 flow is implemented and has produced a successful multi-shot clip with synchronized motion and audio. The reusable Storyboard prompts package that tested flow.
Where it can go: this opens the door to an interactive anime / multiverse experience where users mix characters and worlds, move between roleplay and game modes, make choices, and later talk with characters in the context of recent adventures. That larger experience is the direction of the project; it is not presented here as a finished feature.
I’m also testing how far local video models can go on my own hardware. Seedance-like action quality is a target for experimentation, not a claim of parity.
Code and technical details
100 merged pull requests across Marinara Engine (68) and Agents (32), verified September 5, 2026. My work includes World Maps and spatial state, reusable character/location artwork, storyboard orchestration, host APIs for downloadable agents, recovery paths, and automated validation.
Stack: TypeScript, React, Node.js, ComfyUI, local inference, Playwright, and regression tests.
Daily audio, live voice, and visual input.
Manfred is my wearable AI project. An Omi wearable provides daily audio for transcription and lifelogging; those transcripts can feed a personal LLM wiki that organizes conversations, project notes, and recurring context.
I’m also connecting E09 smart glasses to ChatGPT Live for voice conversations and image uploads. The goal is a wearable interface to the user’s normal ChatGPT experience rather than a blank API session: speak naturally through the glasses, share what you are seeing, and carry useful context back into the personal knowledge system.
Where it stands: voice works through the glasses. Photo capture and delivery have been tested and are still being made more reliable.
This is a prototype. Personal recordings, private context, and credentials are not published.
Technical details
Android/Flutter companion, Bluetooth audio, Python transcription/search services, durable upload buffering, and Tasker phone automation.
Knowledge management plus the local machines behind it.
Demerzel combines daily AI briefs, LLM wiki ingestion from daily conversations, and management of the computers running local AI workloads.
The wiki keeps useful context from disappearing into old chats. Daily briefs surface what changed and what needs attention. Fleet tools monitor supported services, start and stop workloads, and prevent jobs from competing for the same GPU.
Home AI fleet
- AMD Radeon AI PRO R9700
- Larger local language models and MiniMax H3 video experiments, with workload switching when jobs share the GPU.
- Dual NVIDIA RTX 3060 system
- Local language-model serving plus coding and research experiments.
- Single NVIDIA RTX 3060 system
- ComfyUI image-generation workflows.
The goal is to treat the machines as a small shared compute pool that the other projects can use, rather than making every app handle hardware control by itself.
Separate autonomous coding/research workers are still being refined. Wiki content, credentials, recordings, and machine-access details remain private.
Technical details
Python tooling for conversation ingestion, searchable wiki pages, service health checks, workload controls, and explicit boundaries between knowledge, fleet management, and autonomous execution.
School of the Ancients
A conversational learning project where people explore ideas with AI historical mentors. I’ve built web experiences and VR prototypes using React, TypeScript, Node.js, Supabase/Postgres, real-time voice AI, and Unity.
AutomationAI Radar
An automated pipeline that collects AI news and research, filters and summarizes it, and publishes a shorter daily briefing.
Mixed realityAI beyond a chat window
Experiments with AI and digital information in physical spaces, including contributions to a NASA SUITS 2026 student prototype involving assistant workflows, rover-autonomy concepts, and augmented-reality interfaces.
Futurismthetopham.com
A Matrix-inspired home for longer-form ideas about AI, humanity, and what happens if machines become participants in the same economy as people.