I am a specialized AI Systems Engineer focused on building offline-first AI media production tools, self-hosted LLM infrastructure, and multi-agent automation pipelines. My recent work spans the full spectrum — from ONNX model export pipelines and local inference routing to multi-lane video editors, 2D lip sync studios, and LangGraph-orchestrated content automation — all running entirely on local hardware.

• 40+ Active Codebases spanning offline video production, local AI runtimes, voice synthesis, LLM fine-tuning, and enterprise automation.
• Offline-First Philosophy: Every tool I build runs without cloud dependencies — TTS, SFX, video stitching, lip sync, and sketch animation all on-device.
• Production AI Deployments: Veteran of enterprise SaaS backends, browser automation agents, local AI OS builds, and ONNX inference pipelines.

Here is a curated overview of my primary repositories representing deep engineering focus and production capabilities:

A locally-run, offline-first Windows desktop application that automates voiceover generation, sound effect creation, and multi-lane video stitching — entirely without any cloud dependency.

• Engineering Highlights:
  • Multi-Lane Timeline Editor: Drag, reorder, split, merge, and edit clips across 3 independent lanes with a browser-based UI.
  • Local TTS Engine: Generates realistic voiceovers from text scripts entirely on-device using PocketTTS.
  • Local SFX Engine: Generates sound effects from text prompts using ONNX-exported Stable Audio 3 Small, running fully on CPU.
  • Auto-Captions: Word-by-word caption overlays burned into final video via FFmpeg drawtext filters.
  • Background Music Mixer: Mixes ambient music tracks with adjustable volume into the final rendered output.
• Key Stack: Python, FFmpeg, ONNX Runtime, PocketTTS, FastAPI, WebSockets.

A self-hosted, open-source, air-gapped AI runtime Operating System — enabling companies, developers, and researchers to orchestrate state-of-the-art AI agents and multi-model workflows completely locally, from an RTX 2070 laptop to enterprise clusters.

• Engineering Highlights:
  • Universal Inference Layer: Abstracts backend engines (Ollama, vLLM, SGLang, llama.cpp, TensorRT) under a single dynamic routing API.
  • Hybrid Memory System: Merges semantic vector embeddings (Qdrant) and entity-relationship knowledge graphs (FalkorDB) for human-like agent context recall.
  • Distributed Task Orchestration: Celery + Valkey Streams drive parallelized multi-agent loops and complex DAG workflows.
  • Zero API Keys, 100% Privacy: No external phone-homes, no vendor lock-in, fully air-gapped deployment.
• Key Stack: Python, FastAPI, Celery, Valkey Streams, Qdrant, FalkorDB, Ollama, vLLM, Docker.

A locally-run, offline-first Windows desktop application for 2D animators — automatically generating frame-accurate lip-sync video from character sprites and a dialogue script.

• Engineering Highlights:
  • Phoneme Detection: Rhubarb Lip Sync extracts mouth shape timings (phonemes A–X) at frame-level accuracy from generated or provided WAV audio.
  • Character Studio: Upload full-body character images, define face regions, and map custom mouth sprites to each phoneme.
  • Script-Driven Voice: [Character Name] tagged dialogue scripts converted to audio locally via PocketTTS.
  • ONNX Face Tracking: Optional ONNX-based face/gaze tracking for characters with animated heads.
  • Video Compositor: Composites frames + mouth sprites + audio into final MP4s via FFmpeg — no external rendering software needed.
• Key Stack: Python, Rhubarb Lip Sync, ONNX Runtime, FFmpeg, PocketTTS, FastAPI.

A fully automated, sequential content pipeline that discovers viral YouTube videos, writes original scripts, generates TTS audio + AI images + whiteboard sketch animations, and uploads to YouTube — orchestrated by LangGraph multi-agents.

• Engineering Highlights:
  • LangGraph DAG Agents: Full viral_finder → await_approval → data_collection → script_generation → media_generation → upload pipeline with crash-safe resumption via Supabase pipeline_stage state.
  • Thin-State Architecture: LangGraph state holds only IDs — every agent reads inputs fresh from Supabase/GCS and writes outputs back, eliminating context bloat.
  • Human-in-the-Loop Gate: await_approval polls yt_results.approved — a human reviews generated scripts before media production begins.
  • Integrated Sketch Animation: Converts AI-generated images into whiteboard sketch animations using Omni Sketches on GCP VM.
• Key Stack: Python, LangGraph, FastAPI, APScheduler, Supabase, Google Cloud Storage, FFmpeg, Gemini.

An open-source ONNX conversion pipeline for Stability AI's Stable Audio 3 Small — enabling CPU/GPU audio generation inference without a PyTorch runtime dependency.

• Engineering Highlights:
  • Exports the full diffusion pipeline into 4 standalone ONNX modules: Text Encoder → Conditioner → DiT (Diffusion Transformer) → Decoder (Oobleck).
  • Supports both stable-audio-3-small-music and stable-audio-3-small-sfx model variants.
  • Runtime-agnostic deployment on CUDA, DirectML, and CPU via ONNX Runtime — no PyTorch installation required at inference time.
• Key Stack: Python, ONNX Runtime, HuggingFace Diffusers, Stability AI, PyTorch.

A tightly integrated suite of offline voice engineering tools covering real-time TTS streaming, multi-engine workstations, and full voice-driven AI assistants.

• Engineering Highlights:
  • Pocket_TTS: Low-latency real-time TTS streaming via WebSockets (~200ms to first chunk), voice cloning support, CPU-only, glassmorphism web UI.
  • Aura-TTS: Unified offline speech workstation managing engine lifecycles (Kokoro-TTS, Pocket-TTS, Supertonic) via custom mutual exclusion locking under RAM/VRAM constraints.
  • OrpheusAssistant: Offline voice agent with WebRTC bi-directional streams (aiortc), Whisper Large STT, and LLaMA 3.2 3B — integrated with n8n for workflow tool execution.
• Key Stack: Python, FastAPI, WebRTC, WebSockets, PipeCat, PyTorch, n8n, Docker.

A specialized research-to-production pipeline repository detailing high-throughput fine-tuning, reasoning models, and edge optimizations.

• Engineering Highlights:
  • Fine-tuning pipelines utilizing Unsloth kernels for parameter-efficient optimizations (LoRA / QLoRA).
  • RLHF training using Group Relative Policy Optimization (GRPO) without a separate Critic model (pioneered by DeepSeek-R1), targeting Qwen3 8B.
  • Horizontal engineering recipes for Mixture-of-Experts (MoE) kernels and ModernBERT dense classification systems.
• Key Stack: Unsloth, PyTorch, LoRA, QLoRA, HuggingFace Transformers, TRL, JAX.

• 🎬 Offline-First Media Architect: Building complete AI video production stacks that run 100% on local hardware — ONNX inference, FFmpeg rendering, TTS streaming, and lip sync compositing.
• 🧠 AI Pipeline Engineer: Expert in LangGraph DAG orchestration, Kafka event streaming, Celery distributed workers, and multi-agent system design.
• 📦 Docker-first Deployer: Containerizing complex AI stacks with multi-stage builds and clean Docker Compose networking configs.