Evaluation
Compare task results, successful completions and important failures against your requirements.
Task scores and failure examples.
Tools for your portfolio companies
RadianVector tools.
RadianVector tools.
For Venture Capital firms.
Give your AI portfolio companies RadianVector tools. Compare answer quality, response time and cost on company-specific workloads, then keep measuring as models and settings change.
A company can work with us directly, or a venture firm can sponsor the work across its portfolio. When a company needs a new connector, we build it, and it becomes part of RadianVector for every company that follows.
The same RadianVector tools, adapted to the participating company’s workload.
What you getEvidence the engineers can inspect.
Evidence the engineers can inspect.
A baseline that stays current.
Configuration shortlist
Model identifiers, prompts, schemas and supported settings, tied to the conditions under which they were tested.
Task-level comparisons
Correctness, important failures, latency and cost, compared against your baseline.
Validation record
Held-out results and repeatability checks, with the evaluation inputs and scoring needed for future comparisons.
Workload connectors
Connectors that bring each company’s workload into the tools. Every connector we build becomes part of RadianVector.
Ongoing regression checks
The same evaluation rerun on each update, with changes in quality, latency and cost flagged against the baseline.
Built on RadianVectorThe capabilities behind
The capabilities behind
the evaluation.
Use RadianVector tools for workload-specific evaluation and inference measurement, with AI agents orchestrating the experiments.
Workload analysis
Assemble representative tasks, source context and request patterns.
A versioned workload and test plan.
Inference measurement
Inspect response time, throughput, memory use, queueing and runtime behavior under a defined load.
Measurements with their operating conditions.
Configuration testing
Compare models, prompts, output formats and supported serving settings against your requirements.
A shortlist with measured trade-offs.
Verification
Repeat runs, investigate output differences and check the selected configuration on held-out tasks.
A validation record and regression baseline.
inferlint is an open-source component of the RadianVector stack. Explore the engineering and measured examples →
Published engineering workPublished research.
Published research.
Measured results.
Our tools found a bug in mature vLLM code that changed answers when several requests ran at once. We traced the cause, tested a fix and reran the same comparisons to verify the results.
Work with RadianVectorPut your configuration choices
Put your configuration choices
to the test.
Tell us how your portfolio companies use AI today. We will work out with each of them what is worth measuring, then run it with these tools.