AI Product Engineer · Los Angeles, CA
Email · LinkedIn · Repositories
I build software that makes emerging AI useful, from the infrastructure that runs models to the interfaces people use.
Most of my projects start with a practical limitation: a model that will not deploy, a creative workflow that needs too much GPU memory, or two tools that should work together but do not. I work across product design, implementation, and deployment to solve the problem and make the result usable by someone other than me.
I adapt community deployment recipes and patch runtime compatibility gaps to make early-release models available in coding assistants. My development setup includes a two-node NVIDIA DGX Spark cluster and an RTX 3090 workstation.
Deployed GLM-5.3-Flash within 2 days of release and Qwen3.8-Flash-Next within 3 days. In my DeepSeek workload, API gateway optimization raised aggregate concurrent throughput from 199–217 to 275 tokens/s. Supporting work includes shared model APIs, startup automation, health checks, recovery tooling, and qualification tests.
Python · vLLM · Docker · LiteLLM · Linux · OpenAI/Anthropic-compatible APIs
An image-to-3D workflow that brings Tencent's Hunyuan3D into Modly, with textured asset generation, mesh cleanup, and export for game and design workflows.
Reduced measured GPU memory usage from 20.4 GB to 13.0 GB — approximately 36% — at about 5% additional runtime, bringing textured generation within a 16 GB GPU memory budget. Packaged a ready-to-use workflow and prebuilt Windows components, with 300 automated tests supporting repeatable generation.
Python · PyTorch · CUDA · Input/output examples
A desktop application that lets headphone enthusiasts use a two-microphone miniDSP EARS measurement device with Dirac Live, which expects a single microphone input. It handles per-ear calibration and routing, with feedback that helps users spot unreliable measurements.
Fixed clock drift that caused rejected measurements and built 570+ automated tests across installation, measurement, and export workflows. Published Windows and macOS installers.
Public alpha. Windows 11 tested; macOS hardware validation remains pending.
C++17 · JUCE · CMake · GitHub Actions · Website and demo
A desktop extension that puts model settings, session costs, and hardware telemetry in one interface, without requiring changes to the underlying Hermes application.
Built the interface and Python backend, with update-safe installation, rollback, and removal. Added 206 automated tests covering security and compatibility, and submitted upstream improvements for local-model management and routing.
TypeScript · Electron · Python · GitHub Actions
A Claude Code plugin that brings platform-specific design and accessibility guidance into interface development and code review. It loads relevant guidance for the task rather than treating every interface as the same platform.
Published before-and-after examples and a live interface demo so the design changes can be inspected, not just described.
UX systems · Accessibility · Prompt engineering · Claude Code · Examples and live demo
Built and used a review workflow that coordinates agents across pull requests, with repository context, structured tasks, and automated checks before accepting fixes. In one development session, 13 parallel agents resolved approximately 200 code findings across 7 pull requests.
The workflow is in use in my own development. Agent Review Orchestrator, the separate application intended to package it, is currently at scaffold stage.
A private room-correction prototype built around measured acoustic data and a live-tested audio engine. Listener-following behavior currently uses simulated presence; real-room tracking remains research.
My earlier engineering work includes a custom industrial 3D printer, microphone-array design, CAD, and physical simulation.
I use Codex, Claude Code, and Hermes throughout development. I define requirements and constraints, review generated code, and validate changes with automated tests and hardware measurements.
I build supporting infrastructure alongside the product: installers, deployment recipes, health checks, recovery paths, and documentation. I make the distinction between a prototype, a tested workflow, and a released product explicit.
Languages: Python, TypeScript, JavaScript, C++17
AI and infrastructure: PyTorch, CUDA, vLLM, TensorRT-LLM, Docker, Linux, LiteLLM
Applications and delivery: Node.js, Electron, JUCE, CMake, REST APIs, GitHub API, GitHub Actions, Git
AI development: LLM integration, multi-agent orchestration, prompt engineering, Model Context Protocol, inference optimization
I'm interested in AI product engineering, applied AI, and developer tools roles where I can take ownership from prototype through release.
shayanx45@gmail.com · LinkedIn
Performance figures describe measurements from my development setups, not universal hardware benchmarks. Results depend on workload, hardware, and configuration.