Login
AI Platform

Open-source LLMs powering my daily workflow

I run a local-first AI platform across my GB10 and Mac Mini M4 for coding, automation, Discord agents, OCR, tool-calling, and benchmark-driven model selection. This page shows which models I actually use, where they run, how the scheduler routes them, and how I evaluate them under real workloads.

Runtime: Loading…
Platform Focus
Local AI Ops
Inference Runtime
Ollama
Routing Layer
OpenClaw Scheduler
Evaluation Style
Real Workflow Benchmarks

How I use local models

My day-to-day stack is built around local inference, not one-off demos. I use open-source models for coding, multi-step tool-calling, OpenClaw routing, Discord automation, OCR-style extraction, and system workflows running across my homelab.

Instead of treating models like a static registry, I treat them as specialized workers: some are optimized for speed, some for code, some for deeper reasoning, and some for document extraction. The routing layer decides where they run and how traffic gets handled.

AI Platform Architecture

Local Inference Platform

Built a routed local AI platform across GB10 and Mac Mini M4 using Ollama, OpenClaw, and a custom scheduler layer.

Smart Routing + Spillover

Implemented queue-aware host routing, model-class pinning, and Mac spillover to keep interactive workloads responsive.

Benchmark-Driven Selection

Created custom Node.js evaluation harnesses to compare chat, tools, reasoning, and speed under real workflows.

Operational Observability

Exposed live runtime signals into internal dashboards for queue depth, reroutes, spillovers, and current upstream activity.

Models I actually use

Search and filter the stack by purpose. Cards highlight the current live model when the scheduler reports one.

Loading models…

Routing + Scheduler

Checking scheduler

I run a custom Ollama scheduler in front of OpenClaw to make local inference more reliable under real workloads. Interactive traffic queues cleanly, eligible work can spill from the GB10 to the Mac Mini M4, and heavier model classes stay pinned to the GB10.

This turns raw local runtimes into a small routed inference platform with queueing, overflow handling, observability, and benchmark-informed routing decisions.

Scheduler
Loading…

What I optimize for

Most local AI setups break down under real usage: queueing issues, weak tool-calling, or inconsistent routing. My stack is designed to stay useful under actual daily work, not just benchmark screenshots.

  • Tool-calling reliability
  • Multi-step reasoning quality
  • Latency under live load
  • Consistency across repeated runs
  • Practical usefulness in real workflows

How I benchmark models

I benchmark models locally with custom Node.js harnesses that simulate real OpenClaw workloads, not just one-shot prompts. These evals focus on chat quality, tool-calling, agent loops, reasoning, speed, and consistency.

Loading benchmark artifacts…

Why local-first AI

Running models locally gives me control over privacy, speed, cost, and experimentation. More importantly, it lets me build a stack that reflects how I actually work, not how a hosted platform expects me to work.

Runtime philosophy

Every model here earns its place through real use. If it cannot route well, reason well, or stay useful under live workloads, it does not stay in the stack.