Alpha Access and Technical Advisory
I work with emerging AI platforms before they reach the mainstream, testing new models, creative tools, and developer workflows in real production conditions.
This includes client projects, prototype builds, LoRA training, ComfyUI pipelines, and API integrations.
The Thesis
My work sits between research, product feedback, and deployment. I help teams understand how their tools behave in the hands of advanced users, while building an early view of where AI media, agents, and creative tooling are moving.
I choose these collaborations carefully: strong teams, serious products, and tools I believe should exist. The result is a practical R&D loop, early platform access becomes production insight, production use becomes better feedback, and that feedback often turns into open-source tools, workflows, or partner work.
What I Provide
- —Technical feedback on usability, performance, and integration
- —Early bug discovery and edge-case testing
- —Pipeline design for real-time and production use
- —Creative and developer-focused UX insight
- —Open-source tooling, workflows, and implementation notes when useful
Selected Contributions
Selected alpha and partner contributions:
Beta testing → product feedback → AI Partner invitation → Upscale Conference attendee. Continued engagement providing feedback on Magnific and early LoRAs. Invited by Head of Product after practical testing influenced product direction.
Invited to the first FLUX Creators cohort — a small group of creators selected by BFL to collaborate on upcoming FLUX models. Early access to models and features in development, direct line to the engineering team and founders, credits for project use.
Early adopter (2023). Alpha testing of Sound Effects (March 2024). Evaluated for integration into real-time and immersive environments and as sample pack material generation.
Early testing of Glif's LoRA trainer built on Ostris ai-toolkit. Released one of the first LoRAs alongside other artists and engineers during beta. Feedback contributed to platform refinement.
Discord bot testing and long-term community involvement. Contributed PRs and currently focused on Deforum FLUX pipeline development.
MCP server for ComfyUI with Claude Code skill integration. 7 tools for node discovery across 8,400+ nodes, workflow JSON reader, curated packs. Includes Klein/Deforum mappings, LTX-2 workflows. The MCP helps agents know what to do, the CLI does it.
Early API testing and pipeline design for Three.js / R3F — 3D models generated and instantiated directly inside live scenes. Real-time asset generation, not offline workflows.
QA testing across the Comfy ecosystem grew into design and product consultancy. Workflows, edge cases, and production use became both the testing harness and the evidence base for ongoing input on UX patterns, workflow ergonomics, and product direction.
Early access beta testing and workflow development for client projects. Used cloud-based ComfyUI deployments in production workflows.
2B+ tokens used in Cursor alone in H1 2025, experimenting and coding across 15 different models. Compared agentic capabilities through real project work. Provided feedback on Windsurf workflows.
Early testing (first cohort) UI/UX and platform usage. Focused on artist-facing usability and production viability.
Research preview access, max-tier usage. Informs ongoing work on agents using the web browser, agent workflows, and remote development systems.
Open Source Contributions
Alpha testing often leads to open releases:
- •ComfyUI-Koshi-Nodes — Node pack with FLUX nodes, Glitch Candies shaders, dithering. Links outputs across SIDKIT, Glitch Candies, and Flux Motion projects.
- •Koshi-Flux — FLUX-specific ComfyUI nodes for animation and generation workflows.
- •GlitchHunter — LoRA model released during early Glif loradex ai-toolkit testing
- •FLUX-inspired SD 1.5 LoRA model — Enabled animation workflows where FLUX is not AnimateDiff-compatible
- •Deforum PRs — Contributions to core Deforum repository and active development on Flux Motion pipeline bringing latent animation to FLUX.1
- •AI-Utils — Python scripts and notebooks for generative AI: batch image generation, Deforum configs, ComfyUI model installers, RunPod deployment automation
- •UE-AUDIO-MCP — AI-powered game audio toolkit: UE5 plugin + MCP server. Connects MetaSounds to Claude and LLM agents for AI-driven sound design workflows.
- •LMStudio-MCP-Bridge — TypeScript MCP bridge connecting local LLMs to Claude Code for code analysis, documentation, and testing
- •koshi-vox — Voice-to-text recorder for terminal workflows, optimized for Apple Silicon with instant transcription and clipboard integration
- •CreaturesSite — R3F/Three.js landing page with custom GLSL shaders, Rive animations, and AI-generated 3D assets via Tripo3D pipeline
- •Models shared via Civitai and Hugging Face
Continuous loop: early tooling → production use → open release → community feedback.
Why This Matters
This is R&D practice, not a side activity. Beyond platform testing, I write deployment scripts for ComfyUI models and workflow installations, maintain RunPod deployment configurations, and contribute directly to Deforum development. The work connects early access insights with practical tooling — understanding how platforms are built, not just marketed, and translating that into production-ready systems.
Tools
Python • ComfyUI • Bespoke Platform Tools • LoRA Training • API Integration • Pipeline Architecture