Run Quick ML Experiments with AI Agents

Use MCP tools to deploy a GPU, run an experiment, get results, and tear down — all in one conversation. Zero infrastructure left behind.

Run a one-off experiment without any setup overhead. Tell your AI agent what you want to test, and it handles deploy, execute, collect, and cleanup in a single conversation.

---

”Run a quick benchmark comparing PyTorch and JAX matmul performance on an RTX 4090.”

The agent handles everything end-to-end:

1. create_instance — deploys bench-run with RTX 4090
2. instance_status — waits until running
3. ssh_execute — pip install torch jax[cuda12]
4. file_upload — uploads the benchmark script
5. ssh_execute — python /root/benchmark.py
6. delete_instance — terminates immediately

Benchmark results (RTX 4090, 4096x4096 matmul, 1000 iterations):

Framework	Mean time	Std dev	TFLOPS
PyTorch 2.4	2.31 ms	0.04 ms	59.4
JAX 0.4.31	2.28 ms	0.03 ms	60.2

Both frameworks perform nearly identically on this workload. JAX is marginally faster (~1.3%) due to XLA compilation.

**Total cost: $0.12** (20 minutes on RTX 4090 at$0.35/hr). Instance terminated. Nothing left running.

---

”Test if my fine-tuned model loads correctly. Deploy a GPU, load it, run 5 prompts, and shut down.”

The agent runs the validation:

1. create_instance — deploys model-test with RTX 4090
2. instance_status — waits until running
3. ssh_execute — pip install torch transformers
4. file_upload — uploads test_model.py with 5 test prompts
5. ssh_execute — python /root/test_model.py
6. billing_usage — checks cost
7. delete_instance — terminates

Model validation results:

Prompt	Output	Quality
”Summarize this paper…“	3-sentence summary, accurate	pass
”Translate to French…”	Correct translation	pass
”Extract entities from…”	Found 4/4 entities	pass
”Generate SQL for…”	Valid query, correct joins	pass
”Classify sentiment…”	Correct label, 0.94 confidence	pass

All 5 test prompts passed. Model loads in 8.2 seconds, inference at ~45 tok/s. Cost: $0.06 (10 minutes). Instance terminated.

---

”I want to try 3 different LoRA ranks and see which gives the best eval loss. Quick and dirty.”

The agent runs a fast comparison:

1. create_instance — deploys lora-test with A100 80GB
2. instance_status — waits until running
3. ssh_execute — installs dependencies
4. file_upload — uploads training script and small eval dataset
5. ssh_execute — runs with --lora-rank 8 (500 steps)
6. ssh_execute — runs with --lora-rank 16 (500 steps)
7. ssh_execute — runs with --lora-rank 32 (500 steps)
8. delete_instance — terminates

LoRA rank comparison (500 steps each, Llama 8B, 1K sample dataset):

Rank	Trainable params	Eval loss	Train time
8	3.4M	1.842	4 min
16	6.8M	1.687	5 min
32	13.6M	1.691	7 min

Rank 16 wins. It achieves the lowest eval loss — rank 32 doesn’t improve further and trains 40% slower. The extra parameters at rank 32 are likely overfitting on this small dataset.

**Total cost: $0.43** (16 minutes on A100 at$1.60/hr). Instance terminated.

---

Tools used in this workflow

Tool	Purpose
create_instance / instance_status	Provision and wait for GPU
ssh_execute	Install packages, run experiments
file_upload	Transfer experiment scripts
delete_instance	Tear down immediately after results