Create and train your own model with an AI Agent Assistant

StellarFirm is a service where you work with an AI Agent Assistant, the machine learning engineer, to create and train your own model. You describe the model you want. The assistant prepares the data, sets up a reproducible project and proves it on a small run on its own computer. The real training runs on a GPU you control, and any paid compute waits for your Approve. Then it evaluates the model, writes the model card and gets the model ready to serve. StellarFirm assistants do not have GPUs of their own, and no result is guaranteed: the evaluation tells you what you actually got.

Three ways to create and train a model

How it works with StellarFirm

1. 1

   Describe the model

   Tell your CEO assistant what the model should do and share your data. It briefs the machine learning engineer, who asks when something is vague and estimates the GPU memory and hours the job needs.

2. 2

   The assistant sets up the project

   It prepares the data pipeline with leak-free splits and sets up a pinned, GPU-ready project. A small smoke run on a tiny model and a slice of data, on its own CPU computer, proves the code works. It pushes a branch and writes a one page hand-off with the GPU class, the hours and a stop rule.

3. 3

   Real training runs on a GPU you control

   Run it on your own computer, or on Hugging Face Jobs, RunPod, Kaggle or a cloud VM in your own account. On StellarFirm cloud the assistant holds no GPU provider token, so you start the run from the hand-off. Where it can launch compute, that launch waits for your Approve.

4. 4

   Evaluate, document and serve

   The assistant reads the run back, evaluates the model against a baseline, writes a model card and prepares the serving files. Add the served endpoint in Settings, Models, and you choose when to use it.

What skills it uses

The machine learning engineer role carries these skills, written for real training work.

• Dataset preparation

  Cleaning, deduplication, personal data scrubbing, licence checks, splits and checks against your evaluation data.

• LoRA and fine-tuning

  Supervised fine-tuning of open models with LoRA or QLoRA, a smoke run on a tiny model and an evaluation against the base model.

• Preference tuning

  DPO, ORPO or KTO data and training, with checks for reward hacking and length bias.

• Evaluation harness

  Held-out sets, contamination checks, judge rubrics with agreement checks, intervals and regression gates.

• Quantization and export

  GGUF, AWQ, GPTQ, FP8 and ONNX exports, with a measurement of what the smaller model cost in quality.

• Serving

  A typed API or an OpenAI-compatible endpoint with a key, TLS and a load test, plus the steps to connect it in Settings, Models.

• Model card and fairness check

  A model card, checks across slices, the licence chain and limitations, and a private release first.

• Monitoring

  Privacy-aware logging, drift checks, quality against delayed labels and a retraining trigger.

• Tabular models

  scikit-learn, XGBoost and LightGBM with baselines, cross validation, leakage checks and calibration.

• Vision models

  Image classification and detection fine-tunes with per-class metrics and a licence check of the model and the data.

Questions

Does StellarFirm train on its own GPUs?

No. No assistant computer has a GPU. The assistant prepares and proves the project on a small CPU run, and the real training runs on a GPU you control: your own computer, Hugging Face Jobs, RunPod, Kaggle or a cloud VM.

Is my data used?

Your data stays with your work. The training data is prepared in your own project, and the real run happens on the GPU you picked. StellarFirm does not use your data to train its own models.

Can it spend my money?

Never without your Approve. Starting paid compute, such as a GPU job, is a spend that waits for you, and no auto-approve switch lifts it. The assistant gives the GPU class and the hours in the hand-off, and you read the rate on your provider's own page.

Will the model be good?

There is no guarantee. The assistant compares the model with a baseline, reports the numbers and the limits, and says plainly when a result does not hold up. The decision to use or release the model is yours.

Can I use the trained model in StellarFirm?

Yes, on plans that include your own models. Serve it behind an OpenAI-compatible endpoint and add it in Settings, Models. Auto never picks it, so you choose it deliberately.

What does it cost?

Assistant work uses StellarFirm credits, and new accounts start with free credits. GPU time is billed by the provider you choose. This site does not publish prices.