Selected work · Inference & hardware · Agents & embodiment · GitHub activity · AIRewardrop · Connect on X

I'm funboy, founder of AIRewardrop and an independent AI systems builder. I connect agent behavior, developer workflows, inference software and physical infrastructure into systems I can run, inspect and improve.

My path started with gaming, retrogaming and hands-on hardware, then grew into building and running IT and telecommunications businesses: networks, servers, security and customer operations. Crypto and on-chain automation brought another layer: software that interacts with markets and communities. Today, that experience converges in local AI, agent tooling and inference engineering.

I work across the stack: from a conversational interface and its tools to model loading, memory constraints, GPU execution and communication between machines. The connection between those layers is where I do my best work.

Hardware is part of my development process. I build and operate the machines, then work through the constraints that determine whether a model is actually usable: memory capacity, quantization, kernel support, interconnect cost and response latency.

The HaloClu reference deployment runs hybrid W4 GLM inference with tensor parallelism across two nodes, RCCL Socket over USB4 and DFlash2 speculative decoding. The product layer brings that runtime into daily chat and supervised coding workflows.

aireward-llm is the NVIDIA workstation in my lab: 2 × GeForce RTX 3090, an Intel Core i5-13500, 128 GB system RAM and a 2 TB Samsung 990 PRO NVMe. Each RTX 3090 has 24 GB of dedicated GDDR6X memory. It brings agent development and multi-GPU inference research alongside the Strix Halo machines.

My long-term goal is to combine aireward-llm and both Strix Halo nodes into a hybrid local AI cluster. I'm studying how to coordinate dedicated NVIDIA GPUs and AMD unified-memory systems through a common model-selection and orchestration layer.

Local LLM Autopilot / llama.cpp-model-select provides the foundations: hardware-aware GGUF fit planning, CUDA and Vulkan worker selection, model lifecycle control, and recorded performance and quality evaluations. Its cluster design explores independent workers for request routing and replicas, plus ggml RPC for distributed models. Extending those approaches across the mixed hardware is research in progress, guided by memory fit, communication costs and comparisons against local baselines.

Further inference work:

• ds4-multicuda: my fork of antirez's ds4, exploring native CUDA multi-GPU expert placement across consumer GPUs and asymmetric PCIe links.
• StrixHaloClusterDS41: an experimental DeepSeek V4.1 Flash runtime fork of HaloClu, exploring deployment on the same dual-Strix Halo platform.

I keep speed claims attached to their model, quantization, prompt and measurement conditions. Numerical correctness, reproducible tests and retained failure results guide the work. See HaloClu's qualification record for the tested scope and current limits.

AIRewardrop / AIR³ is where my work on agents, interactive products and on-chain systems comes together. I'm interested in the whole interaction loop: what an agent can perceive, what it remembers, which tools it can use and how people stay in control of its actions.

Beyond the projects above, I build the components that give those agents a presence:

• Voice and avatars: Eliza2Face connects local TTS to avatar-ready audio; my Unreal Engine SDK fork explores conversational agents with environment perception and in-world actions.
• Platform integrations: ElizaOS clients for Twitch, Reddit, Farcaster and Telegram.
• Markets and on-chain workflows: AIRTrack for agent trade tracking, RIP2ETF for structured ETF snapshots and ZordBOT for Zcash Ordinal mint orchestration.
• Physical signals and models: Somatic / SomaBridge, a research prototype exploring machine telemetry, learned sensor projections and embodied agent interfaces.

Build across boundaries. Product interfaces, agents, APIs, runtimes and deployment belong in the same engineering conversation.

Make behavior inspectable. Tool activity, memory provenance, telemetry and independent verification help turn a model's output into something a person can evaluate.

Measure on real machines. I use local hardware to investigate memory pressure, numerical behavior and performance, and document the conditions behind each result.

Build with the ecosystem. My work includes original applications, integrations and focused forks. Upstream projects such as Pi, PiLink, llama.cpp and ElizaOS are part of that foundation.

A view of my public repositories and ongoing work, updated daily from GitHub.

A nod to my retrogaming roots, tracing the contribution calendar one square at a time.

Building useful intelligence, from the interface to the machine.
Interested in local AI, agent tooling, inference infrastructure or embodied interfaces?
AIRewardrop · X / @funB0Tnft · Telegram · Explore all repositories

_{Original profile content: 0xfunboy Non-Commercial License · Scope and attribution}