Open research. Real engineering. Upper Michigan.
Exploring local-first AI systems, inference architecture, adaptive computation, intelligent interfaces, and the infrastructure needed to understand and build them.
Superior MI Labs is an early-stage research and development effort based in Upper Michigan.
We are interested in intelligent systems that are understandable enough to inspect, measurable enough to challenge, modular enough to evolve, and useful enough to matter outside a benchmark.
Our work currently centers on AI inference systems, reusable computation, adaptive runtimes, reasoning and decision infrastructure, intelligent interfaces, and local compute.
Adaptive Inference Runtime
AIR is a standalone C++20 inference runtime built around explicit ownership, bounded resources, observable state, and falsifiable optimization.
It currently includes reference/CPU and NVIDIA CUDA execution paths, GGUF model loading within its qualified scope, bounded scheduling and backpressure, generation, runtime observability, a semantic Decision interface, and a local browser command center.
Source and engineering record • Project site • Hugging Face Space
Typed construction substrate for machine-readable systems
Superior MI Builder is the structural substrate for composing, verifying, lowering, simulating, and safely evolving typed systems through one canonical graph authority and one structural mutation path.
R1 completed an eight-wave development and destructive-qualification program covering canonical graph histories, exact artifact identity, deterministic verification, state isolation, explicit effect authority, safe simulation, runtime fault evidence, and a clean-room fresh-agent proof.
Source and R1 engineering record
Inference Fabric explores reusable inference infrastructure, continuity, and ways to reduce unnecessary repeated computation while keeping execution state and system boundaries explicit.
Superior MI Labs is building toward a larger modular research ecosystem.
Areas of active or developing interest include:
- inference architecture and hardware-aware execution
- adaptive runtime systems
- reasoning and decision infrastructure
- reusable computation and continuity
- intelligent interfaces and sensorimotor systems
- local AI and compute infrastructure
- developer and agent tooling
- system observability, deterministic replay, and qualification
- practical applications of advanced AI systems
Not every research effort is public or release-ready. We try to distinguish between ideas under investigation, experimental systems, qualified public software, and mature products.
A few principles show up repeatedly in our work:
Make state explicit.
Important behavior should not disappear into hidden mutable layers.
Keep one authority for each responsibility.
Parallel pipelines and duplicate sources of truth create systems that become
hard to reason about.
Measure before claiming.
Optimizations and architectural ideas should survive testing, comparison, and
falsification.
Design for evolution.
Components should be modular enough to replace, improve, or specialize without
forcing the entire system to be rewritten.
Preserve the evidence.
Failures, negative results, benchmarks, qualification records, and discarded
approaches are part of the research record.
Advanced technical research does not have to be something communities like ours only consume from somewhere else.
Superior MI Labs is being developed with the long-term goal of helping build technical capability in Upper Michigan while contributing useful work to the broader open research community.
Lake Superior and the surrounding region are part of the identity of the lab, not just its visual branding.
GitHub
https://github.com/Superior-MI-Labs
Hugging Face
https://huggingface.co/Superior-Mind-Labs
People • Planet • Progress
Superior MI Labs