Jarad P.

I design and build systems that pair generative AI with production-grade cloud infrastructure. Over 25 years in software engineering and cloud architecture, I've spent the last three years focused intensively on what happens when LLMs move from chat interfaces into autonomous, multi-agent pipelines. My work bridges three layers: model integration (selecting, deploying, and chaining the right models for a task), agentic orchestration (building state machines that coordinate multiple AI agents through plan/write/review/gate cycles), and edge deployment (compiling models to run on-device via WASM, LiteRT, and browser runtimes). I architected a state machine framework (sm3) that forks a single model context into three specialized roles — architect, engineer, and hypervisor — each operating under distinct permission boundaries, producing convergent artifacts without shared read access. This system uses DeepSeek V4 Flash, GPT-5, and Gemini as interchangeable engines within the same orchestration pipeline, with FastAPI dashboards and SQLite-backed persistence tracking every phase run, dispatch log, and sprint. A companion daemon (fw-sm) generalizes this into a DB-driven pipeline where states, transitions, guards, agent assignments, and file contracts are fully data-driven — no hardcoded orchestration logic. On the multimodal side, I have architected video processing workflows: Whisper for speech-to-text, InsightFace for face detection and clustering (DBSCAN), and heuristic-ML hybrid scoring to extract the best subclips. These services ran on GCP with Celery workers, FastAPI/Flask backends, PostgreSQL persistence, and GCS media storage — coordinating through message queues and signed upload URLs. I led the migration of a multilingual social simulation from a Python-server-dependent architecture (llama-cpp GGUF + Argos Translate + Piper TTS) to a fully browser-based stack where Gemma 4 runs via LiteRT-LM WASM for on-device chat and translation, with wllama, WebLLM, and ONNX Runtime Web as fallback runtimes — completely decoupling inference from any backend. The same project involved Piper TTS, Kitten NPZ-based browser TTS, and on-device phonemization. Before going deep on AI, I spent years as a cloud architect specializing in GCP, AWS, and Azure — designing high-traffic, multi-region distributed systems for medical, nonprofit, and digital media clients. I have held certifications across all three platforms and have built microservice backends, media processing pipelines, and full-stack web applications. What sets me apart is the range: contract AI model training and red-teaming, TTS systems from scratch (TensorFlow waveform modeling to browser NPZ engines), Apertium compiled to WASM for browser machine translation, self-hosted LLM inference with llama.cpp, and multi-agent frameworks that decompose engineering work across specialized AI roles. I can work at any level — from designing a cloud network topology to debugging a model's tokenization edge case. I bring this full-stack AI engineering capability — cloud to edge, model to agent, prototype to production — to every engagement. AI / LLM Models GPT-4, GPT-5 / 5.5 Gemini 2.5 Pro Gemini 3.5 Gemini (Gemini CLI) - bootstrap framework Gemma 4 (LiteRT-LM) o3 - Reasoning tasks DeepSeek R1 - Reasoning tasks DeepSeek V4 Flash - Primary experiment model Llama 4 Scout Claude (Claude Code) - CLI agent workflow SmolLM2 (135M, 1.7B) Qwen2.5 0.5B TinyLlama 1.1B Whisper (faster-whisper) Ollama (local) # CLI & Dev Tooling OpenCode CLI - Agent dispatch Google Gemini Agent CLI - CLI bootstrap Codex - CLI agent workflow Bun - Custom tool runtime Docker / Docker Compose Gitea - Self-hosted Git llama.cpp - Local GGUF serving ffmpeg / ffprobe nginx + certbot # Python Libraries & Frameworks Web & API: FastAPI Flask Streamlit uvicorn Jinja2 Mako Bootstrap 5 Authlib Databases & ORM: SQLAlchemy - ORM Alembic - Migrations psycopg (binary) - PostgreSQL driver Prisma (JS) - ORM Task Queue: Celery Redis AI / ML: llama-cpp-python InsightFace tensorflow / Keras argostranslate huggingface-hub transformers torch scikit-learn (DBSCAN) librosa soundfile numpy Game Dev: pygame-ce Kivy / KivyMD Other: python-avatars cairosvg / pillow g2pw / pypinyin / jieba pydantic pytest pycountry # Frontend & Browser Technologies Next.js 16 React + Vite - Frontend TypeScript - Tools & frontends Tailwind CSS shadcn/ui TanStack React Query NextAuth.js ESLint / Prettier Browser AI Runtimes: phonemizer # Edge & On-Device ML Gemma 4 LiteRT LiteRT / ai_edge_litert WASM Inference - LiteRT-LM, wllama, WebLLM, ONNX Runtime Web Apertium WASM Pygbag WASM Pyodide Piper TTS (ONNX) Kitten TTS