AXIS
A deterministic load router for AI and compute infrastructure
Hi, I’m Marcus, founder of Augustus.
I build systems that make AI infrastructure faster, leaner, and more predictable.
Where Chariot routes models, Axis routes compute deterministically.
Over the past weeks, I’ve been developing Axis Engine,
a load-balancing core built for the new generation of AI workloads,
where every millisecond and every GPU counts.
🧠 Why I built it
Modern AI systems distribute billions of micro-tasks per second: embeddings, tokens, inferences, gradients.
Most of that routing still relies on probabilistic hashing or ring-based balancing.
It works, but it’s noisy. Nodes shift, caches reset, GPUs idle.
I wanted to build something exact:
a system that routes with mathematical precision, not chance.
Axis doesn’t balance traffic, it orchestrates it.
Each request is mapped to its destination through a single deterministic formula:
integer-pure, reproducible, and mathematically exact.
No floating-point drift, no random seeds, no rehashing.
⚙️ How it works
Axis replaces randomness with mathematical structure. Each request is evaluated through a deterministic scoring process that ensures every node receives its exact share of work.
Instead of relying on probabilistic rings or random hashing, Axis calculates precise mappings that remain stable even as the system scales or changes.
The result is a routing engine that maintains perfect balance under dynamic conditions fast, predictable, and mathematically fair.
📊 Benchmarks
Tests on real hardware (8 threads, K = 1024):
| Metric | Result |
|---|---|
| Throughput | 139 M ops/sec |
| Latency | 7.18 ns/route |
| Churn (K→K−1) | 0.095 % |
| Determinism | 0 mismatches / 1 M keys |
| Max deviation | 3–5 % |
That’s roughly 2–3× faster than traditional consistent-hashing engines
with 100× less churn during node changes.
In AI inference pipelines, that means:
- Less GPU idling
- Fewer cache resets
- Lower routing overhead
- Predictable, repeatable scaling
🧩 Current stage
Axis is in development (TRL 6), a fully functional prototype proven on real hardware.
Next steps include long-term soak tests, observability endpoints, and pilot integrations with AI-infra partners.
If you operate large-scale inference, model serving, or distributed training systems,
you’re exactly who Axis is built for.
If you’re working on similar infrastructure challenges or want to experiment with deterministic routing,
I’d love to connect and exchange ideas.