Products

Find what your workload can safely do without.

Three scoped offerings, supported by one evidence and enforcement foundation. Each engagement measures whether savings exist on your system.

Private pilot

01 / Measure and certify

Optimization Profiler + Scoped Certification

Find the cheapest configuration that actually survives your workload.

The Optimization Profiler compares baseline and optimized configurations on a frozen deployment identity. It measures the complete cost of profiling, policy selection, execution, and fallback; checks quality across workload groups; calculates finite-sample failure-risk bounds; and produces an auditable approval, rejection, or insufficient-evidence result.

What you receive

Measured memory, latency, throughput, training-work, and cost changes, with baseline-versus-candidate comparisons.
Subgroup and workload-cluster results, observed failure counts, and statistical upper bounds.
An immutable evidence bundle and an exact deployment scope.
A deployable certificate, an explicit rejection, or an insufficient-evidence decision, with recertification and fallback requirements.

Availability: The integrated harness supports measurement, evaluation and evidence inspection. Public runs are restricted to configured presets and bounded capacity; paid evaluations are scoped to the customer workload.

Validated pilot

02 / Serving memory

KV Optimization Calibrated per model

Reduce model-serving memory with a policy calibrated to the model and workload being deployed.

The KV Optimization service evaluates fixed and adaptive cache-retention policies, measures their physical memory use and quality, and identifies the smallest policy that remains inside your defined quality and risk envelope. Certified policies can increase concurrency and reduce cache pressure, while unsupported or changed deployments fall back to the full cache.

What you receive

Offline KV qualification and a fixed-recipe recommendation.
An optimization report and deployment manifest.
A private integration pilot, with per-model calibration for adaptive policies.

Availability: The KV-compression endpoint and per-model router are implemented. Supported model, backend and policy identities remain exact; live public comparison availability depends on the configured preset and operator-owned capacity.

Read the scoped memory results

Research pilot

03 / Repeated adaptation

Training Optimization Bounded adaptation

Reduce repeated adaptation work while separating nominal work reduction from actual runtime savings.

The Training Optimization service analyzes fine-tuning and adapter-training workloads, identifies examples or updates that may be omitted, and measures the resulting quality, backward work, wall time, and total cost. It recommends the simplest successful policy rather than assuming that a more sophisticated selector is automatically better.

What you receive

An offline training-efficiency assessment or repeated-adaptation pilot.
A recommendation or training plan for a defined workload.
A scoped research engagement that measures end-to-end economics.

Availability: The owner-confirmed training endpoint and integrated harness support scoped evaluation. Completed ML-T6R measurements are inspectable in the demo; public live runs require the existing optimized runner contract and a bounded compute allocation.

Read the work and wall-time results

What our availability labels mean

Integrated platform foundation: Implemented in the control or enforcement layer.
Private pilot: Customer engagements are possible; the workflow is not general self-service.
Validated pilot: A scoped technical result exists; deployment must be reproduced on the customer’s stack.
Research pilot: The mechanism is promising; modern production economics are not yet established.
Research tool: Useful implemented research software, separate from the commercial runtime.
In development: Design or implementation work exists; completion is not verified.

Scoped evidence

The result and its limits, together.

These are registered experimental results. They do not authorize a different model, workload, backend, or hardware configuration. Your workload still needs its own measurement.

Fixed KV policy · Qwen3-8B

7.877× fewer KV-cache bytes

On one frozen Qwen3-8B evaluation, a 12.5% KV-retention policy allocated approximately 7.877× fewer KV-cache bytes—about 87.3% less KV memory—while remaining within the registered quality envelope.

This measures KV memory. It does not establish an equivalent latency or cost reduction. Results are model-, workload-, backend-, and hardware-specific.

Adaptive KV policy · Per-model calibration

28.7–35.7% fewer KV bytes

The per-model adaptive router used approximately 28.7% to 35.7% fewer KV bytes than the matched-risk fixed policy on its registered panels, with zero observed underallocations across 495 evaluated requests.

Cross-family transfer failed. Each router requires calibration to its own model and workload. Zero observed failures is not a guarantee of zero risk.

ML-T6 · Bounded adaptation

50% fewer backward examples

The evaluated adaptation pool fell from 4,800 to 2,400 backward examples, with zero observed failures in 149 certification evaluations and a one-sided 95% failure-risk upper bound of 1.99048%.

The registered fully charged wall-time reduction was 1.69%. This does not establish a 50% reduction in training time or cost. End-to-end savings must be measured separately on each training system.

About the pre-cache follow-up

A later pre-cache implementation reported a larger CPU wall-time improvement. It remains follow-up engineering evidence: cache-construction cost and amortization across repeated jobs must be evaluated before drawing a production cost conclusion.

The historical demo below replays a separate, frozen 3.94× fewer resident KV-cache bytes result at a 25% cache budget on a correlation-broken panel, using Qwen3-8B on Apple M5 Pro, PyTorch MPS, bf16, and SDPA. It is a different evaluation from the 12.5% retention result above. The endpoint does not run Qwen live.

Run the historical memory comparison

Executable API exampleCompare measured resource use

Live FastAPI gateway

POST /v1/demo/resource-comparison

import json
from pyodide.http import pyfetch

response = await pyfetch(
    "https://api.reductionofstates.com/v1/demo/resource-comparison",
    method="POST",
    headers={"Content-Type": "application/json"},
    body=json.dumps({"scenario": "qwen3-8b-kv-cache"}),
    credentials="omit",
)
response.raise_for_status()
comparison = await response.json()
comparison_json = json.dumps(comparison, indent=2)
print(comparison_json)
comparison_json

Pyodide · runs in this browser

const response = await fetch(
  "https://api.reductionofstates.com/v1/demo/resource-comparison",
  {
    method: "POST",
    headers: { "Content-Type": "application/json" },
    body: JSON.stringify({
      scenario: "qwen3-8b-kv-cache",
    }),
    credentials: "omit",
  },
);

if (!response.ok) throw new Error(`Gateway returned ${response.status}`);
const comparison = await response.json();
console.table(comparison);

JavaScript · runs in this browser

curl https://api.reductionofstates.com/v1/demo/resource-comparison \
  -H "Content-Type: application/json" \
  -d '{"scenario":"qwen3-8b-kv-cache"}'

No API key required

Request body Sent directly from your browser when you run the Python or JavaScript example. Edit the JSON to test endpoint validation.

Measured comparisonNaive execution vs Reduction of States

Not run yet

Naive full-cache execution—

Reduction of States—

Resident KV bytes saved — Run Python or JavaScript to load the measured result.

Illustrative cost calculator

Calculate your potential AI savings.

Use a measured cost reduction after charging profiling, policy selection, execution, fallback, and any cache construction. Memory reduction or fewer backward examples alone cannot establish dollar savings. The M5 Pro evidence is not an input to this calculator and does not establish NVIDIA savings.

Your inputs

Build a savings scenario

The example uses assumed cost inputs. Replace them with measurements from your actual backend and hardware.

Current monthly AI infrastructure spend GPU cloud charges, owned-infrastructure cost, or both. Share of monthly spend covered by the measurement The share of baseline cost attributable to the measured workload—not its share of requests. Fully charged cost reduction on that spend Include all overheads. Do not enter a memory-only or example-count reduction.

Estimated impact

Potential annual savings

$60,000

Monthly savings: $5,000
Projected monthly spend: $45,000
Effective reduction: 10%

Current spend10% potentially avoidable

With 40% spend coverage and a 25% measured reduction, the effective reduction is 10%.

Formula: monthly spend × covered spend fraction × measured cost reduction. This scenario does not measure serving capacity or throughput.

Put your workload to the test.

Discuss a scoped assessment and find out which configurations are worth deploying.

Discuss a private pilot