Projects
What I've built.
Each project below is something I built and deployed. These aren't tutorials or demos—they're working systems running in production or deployed to real infrastructure.
AssetFlow
Flagship AI production platform
End-to-end production AI platform that orchestrates content pipelines from creation through narration, rendering, and delivery. Manages assets across stages with automated handoffs, human-in-the-loop quality review, and full pipeline visibility. Built across three major versions, each expanding capability.
- Multi-stage pipeline orchestration with status tracking and automated handoffs
- AI-powered content generation: article writing, narration, image creation, video rendering
- Event-driven asset ingestion from cloud storage with version awareness
- Human-in-the-loop review gates for quality control
- RAG system using Gemini File Search grounding with corpus management and citation extraction
- Multi-user access control with per-user workflows on shared asset libraries
- YouTube integration with OAuth2 for automated publishing
F5-TTS Voice Cloning
Custom voice synthesis on serverless GPU
Fine-tuned the F5-TTS model on ~2 hours of custom voice recordings for production-quality voice cloning. Deployed on serverless GPU infrastructure with a custom Docker container, CUDA 12.1, and an API supporting multiple output formats. Discovered and documented critical, previously undocumented reference audio requirements that eliminate common synthesis artifacts.
Qwen3-TTS Voice Cloning
1.7B-parameter model on serverless GPU
Deployed Qwen3-TTS (1.7 billion parameters) on serverless GPU as an additional voice cloning endpoint. Uses flash attention for efficient inference with S3 storage for generated audio.
AI Video Generation
Text-to-video & image-to-video at 720p
Built and deployed the Wan2.2-TI2V-5B diffusion model (5 billion parameters) on serverless GPU for text-to-video and image-to-video generation. Produces 720p video at 24fps in landscape or portrait, with configurable duration (2–5 seconds), guidance scale, and seed control. Supports batch processing and optional S3 storage.
BookForge
AI-powered book creation for Amazon KDP
End-to-end non-fiction book creation pipeline. Two-phase workflow: AI-driven market research and niche discovery, then chapter-by-chapter generation with citations and KDP-formatted DOCX export. Uses Claude for writing and Perplexity for real-time research, with a structured service layer pattern for each stage.
SlideFlow
AI presentation generation from articles
Converts long-form articles into structured slide decks using a strict 11-type JSON schema. Generates content slides, scripture layouts, key points, two-column comparisons, and image placements—each with speaker notes. Configurable image frequency, detail levels, and audience targeting. Outputs to PowerPoint via python-pptx.
NLP Translation Pipeline
30,000+ verses with morphological analysis
AI-assisted word-by-word translation pipeline processing 30,286+ verses across 66 Biblical books (Hebrew OT and Greek NT) with full morphological analysis. Processed 444,785+ individual words with morphological tagging, etymology, Strong's cross-references, and interlinear formatting. Built a structured lexicon database with 475K+ SEO-friendly pages and etymological comparisons across 15 Niger-Congo Bantu languages.
Pocket TTS Integration
CPU-based TTS at 6x real-time
Integrated Kyutai Labs' 100M-parameter Pocket TTS model as a lightweight, CPU-only text-to-speech option for AssetFlow. Runs at 6x faster than real-time on consumer hardware using only 2 CPU cores, with ~200ms latency to first audio chunk. Supports voice cloning and handles unlimited-length text input.
See the code.
Browse my public repositories on GitHub or get in touch to discuss a project.