AI runtime engineer / agent systems builder / human-machine interface designer.

I build AI systems from the runtime up: MLX and Metal model ports, local inference services, agent harnesses, memory loops, world-model experiments, and interfaces that make AI observable, steerable, and trustworthy.

Current direction: practical AGI scaffolding, Thinking Machines, local-first agents, world models, and HMI for both humans and AI.

Portfolio / Technical resume / GitHub / [email protected]

_{Portfolio identity / selected systems / AI runtime skill map}

Some current Meadow and agent-runtime work is intentionally private or staged. Public repos show sanitized toolkits, benchmarks, demos, and portfolio evidence. I do not publish unreleased code, private credentials, client material, or unreviewed internal experiments.

cuda2mlx turns repeated PyTorch/CUDA to Apple Silicon MLX porting work into an analyzable framework.

• Covers LLM, vision, 3D, sparse ops, signal workloads, and custom Metal kernel paths.
• Includes state_dict rename, tensor layout conversion, cookbook ops, hard-case markers, and parity gates.
• Public README documents 11/11 E2E parity tests, 37/37 smoke coverage, and real repo analysis on TRELLIS, ProPainter, LaMa, InstantMesh, and LGM.

I work on the layer between model capability and product reliability: runtime latency, memory pressure, cache behavior, streaming, and service lifecycle.

• mlx-qwen3-tts-real-streaming: chunk-level streaming path for Qwen3-TTS on MLX.
• mlx-audio: STT/TTS/STS library work around Apple Silicon speech analysis.
• mlx-sci: scientific and signal-processing kernels on MLX.

vsmonster explores a local agent operating system for software work.

• One chat interface orchestrates local/cloud AI agents.
• UFO plans and queues tasks, BlueMonster executes, and messaging bridges route work from phone to desktop.
• The product direction is not just "AI chat in an editor"; it is task lifecycle, review, recovery, and reusable skill accumulation.

AI_shader_Music_lab is a browser-based generative instrument.

• WebGL shader core, WebAudio rhythm engine, and p5-style generative panels.
• 21-cell audio-visual dashboard with bidirectional coupling: rhythm drives visuals, and visual metrics feed back into sound.
• This work sits in the HMI line: interfaces that make system state felt, not hidden.

• brainwave-eeg-interface: EEG interface experiments with real-time FFT and GPU rendering.
• emg-gesture-recognition: 8-channel EMG gesture recognition.
• ai-visual-web-inspector: visual web inspection and UI decomposition with browser automation.

The public repositories are organized into four layers:

1. Core AI systems: MLX, Metal, inference, local serving, speech, and parameter-constrained training.
2. Agent harnesses: task routing, review, verification, memory, and tool execution.
3. World and interface demos: 3D, visual state, generative media, HMI, and embodied probes.
4. Archive / older product prototypes: useful as breadth evidence, but not the center of my current AI direction.

For the most complete and current story, start with the portfolio:

• fawstudio.com/shengkai
• fawstudio.com/portfolio

MLX / Metal / Apple Silicon / PyTorch / CUDA porting / state_dict conversion / tensor layout / custom kernels / parity tests / local inference / KV cache / chunked prefill / TTFT / streaming TTS / agent harness / tool use / memory verifier / world model / 3DGS / HMI / WebGL / Three.js

I started from physics and spent years in creative technology, visual systems, and interface design before moving deeper into AI runtimes and agent systems. That mix matters: I care about whether AI is fast, measurable, controllable, and also understandable to the person using it.

My current fit is strongest where product, runtime, and interface meet: AI systems that need to be built, profiled, debugged, and made usable.