Resume — Michael Siarkowski

Proof of Work

Designed and shipped Pentecost, an AI dubbing platform (diarization → LLM translation → voice-preserving TTS with voiceprint verification → multitrack review console); put into production in September 2026 for studio staff and volunteer linguists, 7,500+ lines dubbed across 19 videos in six languages; voice drift cut from 19.7% to 0.5%.
Built V, a multi-AI collaboration system that runs like a small company: I speak to a voice assistant, AI project managers write the briefs, AI departments deliver against evidence, and nothing is pushed, deployed, or sent without my spoken OK. Departments hand work to each other on their own, and an audit department reviews how the whole system runs and turns repeated mistakes into new rules for my approval; 60+ tickets closed across 5 projects since October 1, 2026; 1,300+ automated tests.
Designed, coded, and shipped mico, a desktop AI production toolbox with a storyboard pipeline and per-project cost tracking (1,500+ automated tests); by our own count it roughly halved pre-production time.
Generated ~30 minutes of AI B-roll, solo, for a ~90-minute documentary in ~4 weeks at ~$3K generation spend; the film drew 100K+ YouTube views in its first week.
Sole 3D artist owning a complete UE5 pipeline; built two 5-camera stages with real-time on-stage graphics; promoted from editor to department lead in four years.

A media production studio producing documentaries, commercials, and original programming; runs a 250K+ subscriber YouTube channel. Tennessee.

Lead, VFX & AI2025 – Present

Senior VFX Artist2023 – 2025

VFX Artist2022 – 2023

Video Editor2021 – 2022

AI Tools I Built and Shipped

Pentecost (Python, FastAPI, React/TypeScript, WhisperX, pyannote, Demucs, Qwen3-TTS, llama.cpp, Cloudflare Tunnel): turns an English video into six dubbed languages with every speaker's own voice; three-pass LLM translation under per-line duration budgets; naturalness gate (UTMOS) plus voiceprint, gender, and duration verification with automatic retry; browser review console with one timeline lane per speaker, per-line history, presence, roles and invite links. A 17-minute 4K video dubs in ~26 minutes per language on one workstation. Built solo in three weeks; put into production in September 2026 for studio staff and volunteer linguists.
mico (Tauri v2, React/TypeScript, Rust commands; Claude, gpt-image-2.5, Seedance APIs): a desktop production toolbox — deterministic prompt tools, image and video generation with a cost model derived from vendor price pages, project folders, drag-and-drop assets, R2-hosted reference video, a 360° panorama tool, and a storyboard pipeline (idea → concept → script → shot list → pre-production cards → panel grid, manual or one-click). 1,500+ automated tests, spec-driven changelog across three product generations. Used by the studio team.

Generative AI Production

Video generation on Seedance, chosen for the quality of its output; image generation with gpt-image and Stable Diffusion/ComfyUI (custom LoRAs); voice cloning (GPT-SoVITS, Qwen3-TTS); LLM integration and agentic coding (Claude, GPT, local models via Ollama/llama.cpp). Generative AI in production since 2022.

Real-Time 3D / VFX

Own the studio's entire 3D pipeline as its sole 3D artist — modeling, look-dev, environments, rigging, animation, motion capture, FX, rendering, compositing — across Unreal Engine 5, Houdini, Nuke, Blender, ZBrush, Marvelous Designer.
Build real-time cinematics in UE5 — Sequencer, Blueprint, Niagara, MetaHuman, Chaos, Lumen/Nanite — for documentary and commercial productions, including 3D for a second documentary (unreleased).
Promoted to department lead in 2025; hired and mentored a VFX artist; shipped 80+ videos and 5 major productions.

Studio Engineering

Designed and built two 5-camera recording stages end-to-end — cameras, lighting, audio, encoding workstations, streaming — and run real-time UE5 on-stage graphics with virtual-production workflows.

Video Editor2019 – 2021

Edited event and promotional videos for local clients; designed and installed livestream setups.

V — Independent project: a voice assistant on macOS that coordinates my Claude Code sessions. Its memory grows with one user: a skill book where a new habit stays on trial until it works three times, so corrections become rules, with no model retraining. Local speech and voiceprint on MLX.
Mecoree (mecoree.com) — Portfolio as a work sample: raymarched light scene and liquid-glass card material on raw WebGL, spring-physics carousel, fixed pixel and frame budget; built from my own mockups with AI-assisted coding in two days.
Client web platform — Bilingual (EN/ZH) website and self-service back office for a 100-member organization: Next.js 16, Supabase with row-level security, four staff roles, draft-to-publish content, YouTube sync cron, first-party analytics, Lighthouse desktop 100 / accessibility 94–96; spec-driven delivery with two pre-launch audits. Delivered, launch pending.
AI-directed character lineup — Consistent hand-painted-3D style across a cast of game characters and creatures, generated and art-directed with image models.

Generative AI

Text-to-Video · Image-to-Video (Seedance)
Image Generation (gpt-image, Midjourney, SD/ComfyUI, LoRA training)
Character & Style Consistency
Voice Cloning & AI Dubbing (Qwen3-TTS, GPT-SoVITS, ElevenLabs)
Speaker Diarization (WhisperX, pyannote)
LLM Integration · Agent Orchestration (Claude Code hooks, multi-session workflows) · Local Models (Ollama/llama.cpp)
Prompt Systems · Cost Modeling

Real-Time 3D / VFX

Development

Editorial / Post

Client & Production

Brief-to-Delivery with Revision Rounds · Pitch Decks · Client Reviews on Live Previews
Rapid Prototyping · Workshops & Enablement for Non-Technical Teams
Multi-Camera Stage Buildout · Portrait Photography

Rensselaer Polytechnic Institute (RPI) — Troy, NY2019

B.S., Electronic Media, Arts, and Communication (EMAC)

◆