Optimization that fits your development workflow.

Run Artemis from your editor, during code review, before a release, or on a schedule. Every run ends with a pull request containing the best-performing version and benchmark results.

Let your coding agent write the code.Let Artemis optimize it.

When a coding agent reaches an optimization task, it hands it to Artemis. Artemis explores multiple solutions, benchmarks them on your workload, and returns the best-performing one. Continue working in your terminal while Artemis searches in the background.

Command Console

live

ISS-6 · Attention falls back to the slow kernel

objectives · latency · memory · same output

steered from your agent

discoverwaiting

Nothing running. A discovery starts when an issue is picked up.

ISS-6

EXP-FDB88%
Expand K/V to a stride-0 view
qwen3-coder
EXP-9DF74%
Materialize K/V per beam
gpt-oss-120b
⋯

v4+0.57
expand() before permute
devstral
⋯
v10+0.63
guarded on the batch mismatch
qwen3-coder

steering · your reasoning modelcoding · qwen3-coder · gpt-oss-120b · devstral

PR ready · #318v10 · qwen3-coder

cross-attn call4.9× faster

424µs

87µs

real-time factor−16%

0.1349

0.1131

discoverwaiting

Nothing running. A discovery starts when an issue is picked up.

ISS-6

EXP-FDB88%
Expand K/V to a stride-0 view
qwen3-coder
EXP-9DF74%
Materialize K/V per beam
gpt-oss-120b

v4+0.57
expand() before permute
devstral
v10+0.63
guarded on the batch mismatch
qwen3-coder

steering · your reasoning modelcoding · qwen3-coder · gpt-oss-120b · devstral

PR ready · #318v10 · qwen3-coder

whisper/model.py+7 −0

@@ -1021,7 +1021,7 @@ def cross_attn(q, k, v, mask=None):

−scores = torch.einsum('bhdn,bhdm->bhdnm', q, k)

+scores = flash_attn(q, k, v, causal=False)

1022attn = torch.softmax(scores, dim=-1)

1023return attn @ v

cross-attn call speedup

4.9× faster

Winner (v10)

87µs

Baseline

424µs

real-time factor improvement

−16%

Winner (v10)

0.1131

Baseline

0.1349

Turn every code changeinto a better one.

Artemis reviews the diff the way a performance engineer would, then returns a branch with the optimization already applied and benchmark results attached.

github.com/your-org/speech-stack/pull/318

opened by artemis

openartemis/attention-fast-path → main

Keep attention on the fast kernel

Measured on your runner

	Baseline	After
Attention call	424 µs	87 µs	4.9×
Real-time factor	0.1349	0.1131	−16%
Accuracy	2.070%	2.070%	unchanged

benchmark
equivalence
tests

Scan the whole codebase,not just the diff.

Artemis scans every module in your codebase, validates each optimization on your workload, and opens a separate pull request for every improvement that performs better.

sweep

scanning

release candidate0/5 modules

whisper/—
whisper/normalizers/—
tests/—
notebooks/—
scripts/—

0
files: 0
issues: 0
discoveries

pull requests back0 open

#318−16%
Keep attention on the fused kernel
#3191.4×
Cache the filterbank across calls
#320−18%
Index the segment list

Nothing raised yet. A PR opens once a module's fix is benchmarked.

Give it a schedule.Wake up to pull requests.

Run Artemis on your own infrastructure, on your own schedule. Queue as many optimization tasks as you like and wake up to validated pull requests.

schedule

every night

a local model$0 api spend

one issue a night0 merged

+4.1%Mon
+2.6%Tue
+1.9%Wed
+3.4%Thu
+2.2%Fri
+1.4%Sat
+2.8%Sun

Each night Artemis takes the highest-impact issue it can prove a win on, benchmarks it, and leaves a PR waiting for you.

queue · by impact

ISS-692
Attention falls back to the slow kernel
ISS-978
Redundant allocation in log-mel
ISS-764
Repeated dict lookup in the token loop

Discover the ROI hiding in your stack

Point Artemis at a system you already run, and see the improvement it finds, validated, before you change a thing.

Book a demo