Skip to content
Developer experience

GPU performance you can measure.
Speedups you can trust.

A CLI for your terminal, an extension for your editor, and a gate for your pipeline — all backed by the same verification engine, so the number never changes as it moves from your machine to main.

Request early access Explore the platform
Built for the terminal

Fast, scriptable, and honest by default.

Kernalin lives where you already work. Every surface is thin over one engine, every result is greppable, and nothing ships a speedup it can’t prove. Below: the CLI, the editor, and the PR gate.
Available
The rnx CLI

One command. The whole lifecycle.

rnx is the Kernalin CLI, a Python package. Ingest a workload, profile it, optimize it, and gate it — each a subcommand of rnx. Output is line-oriented and prefixed with RNX_* keys, so you can pipe it straight into grep, jq, or a CI step without parsing prose.
rnx verify Run the 5-stage correctness gate and time a candidate against your oracle; add --baseline for a statistical verdict.
rnx profile-run Record a kernel profiling session into run history. Hardware counters (occupancy, bandwidth, tensor cores) need Nsight counter access.
rnx memory Analyse recorded memory stats for fragmentation, allocation churn and uncoalesced access, with recommended fixes.
rnx optimize Ask a model (Claude by default) for rewrites; only candidates that pass correctness and anti-cheat, and are measurably faster, are kept.
rnx gate Compare a candidate to a stored baseline and emit the BLOCK / WARN / INFO / PASS tier a CI job acts on.
rnx ingest Discover kernels and run intake checks on a source tree or file before anything executes.
zsh · rnx 
$ rnx ingest kernels/softmax.py
# discovers kernels and runs intake checks
$ rnx verify kernels/softmax.py --oracle kernels/softmax_oracle.py
RNX_STATUS=…  RNX_CORRECTNESS=PASS|FAIL  RNX_MAX_ABS=…
RNX_CANDIDATE_MEDIAN_MS=…  RNX_ORACLE_MEDIAN_MS=…
$ rnx optimize kernels/softmax.py --oracle kernels/softmax_oracle.py
# model-proposed rewrites; only verified, faster ones are kept
$ rnx gate kernels/softmax.py --oracle kernels/softmax_oracle.py \
    --baseline-store .rnx/baselines
RNX_GATE_TIER=BLOCK|WARN|INFO|PASS  # CI acts on the tier

Real commands and flags; output abbreviated to the keys each command prints. The CLI is available to early-access users.

Preview Not yet on the VS Code Marketplace; available to early-access users. It is a thin client — the engine does the work.
VS Code extension

The profiler’s verdict, in your gutter.

See which functions the engine can act on, read their metrics inline, and run verify, fix or optimize without leaving the file you’re in.

Optimizable functions, flagged

The extension marks the kernels and hot functions the engine can act on, so you spend attention where a rewrite is actually available.

Inline metrics

Latency, percent of peak, and whether the kernel is compute- or memory-bound appear beside the code, from the engine’s last run.

Verify, fix and optimize from the editor

Run the engine’s verify, fix and optimize commands on the current kernel. Fix shows its proposed diff; optimize reports the verified result and its run log.

Same engine underneath

Every action calls the contracts-first engine the CLI and CI use — so the editor never shows a number the pipeline later disputes.

Every surface Kernalin plugs into
Available
CI/CD · GitHub Action

A regression can’t merge in silence.

Drop the Kernalin Action into a workflow and every pull request gets a performance gate: baseline versus candidate, the same correctness / anti-cheat / regression checks the CLI runs, and a verdict posted where review already happens. Green means proven — credibly faster and correctness-clean — not merely different.
How the verdict is decided
github.com · pull/482 · checks
Kernalin Performance Gate bot
Baseline
2.84 ms
Current
1.91 ms
Change
−32.7%
Correctness PASS
Anti-cheat PASS
Regression PASS
Verdict Pass · merge allowed
By design

Minimal JS. Fast by default.

The tools stay out of your way. The CLI is a pip-installable Python package with line-oriented output, the extension is a thin client that offloads work to the engine, and the CI gate adds one step, not a build system. Nothing here asks you to change how you already ship.
Request early access Explore the platform
Early access is open

Measure it. Prove it. Ship it.

Bring the rnx CLI, the VS Code extension, and the CI gate to your GPU workloads. Join the early-access program.

Request early access Explore the platform