PROBLEM IN, MODEL OUT.

For the thinking, yes: scoping the problem, labeling data, writing examples and writing the code for custom tasks all run on your own subscription, in an isolated session. Engram supplies the compute, the evaluation and the hosting, so there is no markup on the intelligence. Without an account you can still train on a labeled upload or a public dataset with a deterministic planner.

It is a private beta and free for now: you bring the AI account, we cover the compute. Either way you see the price before you spend it: a built-in estimate in the dashboard, your agent's own estimate for custom jobs, a monthly cap, and a max-cost setting that refuses to start a job whose estimate is higher. Web search for datasets is capped per job.

Uploads are stored on Engram's cloud storage, and training, labeling and dataset building run on cloud machines. Planning and labeling go through your own AI account. Nothing is shared between users, and custom code runs in a sandbox that cannot see other people's data. If your data cannot leave your perimeter at inference time, download the bundle and run it yourself: it is weights, code and a Dockerfile.

Yes. Every model comes with a downloadable bundle: the weights, the prediction code, a server, a Dockerfile and a model card. Nothing is locked in. A live endpoint is there too, and it scales to zero when nobody is calling it, so quiet hours cost nothing.

The test set is split off before any design is chosen and is never used for training. Duplicates are removed before the split, a baseline sits next to every number, and you get a confidence interval and a verdict that is allowed to say no: delivered, below your bar, or infeasible. For custom tasks, Engram runs your agent's predictor on the held-out inputs and scores the outputs itself.

Classifiers for text and tables, image and audio classifiers, text-to-text models, fine-tuned open language models, and anything else you can show with examples (transcription, detection, reasoning with checkable answers) through custom tasks where your agent writes the code. The GPU options run from a T4 up to eight H100s.

Yes, as a custom task: your agent writes a LoRA or QLoRA fine-tune of an open model, picks the GPU, and gives you a time and cost estimate before training starts. Sandboxes last at most 23 hours. We have run the full path on small models; very large ones use the same route with bigger GPUs, so read the estimate and set a max cost.

Build it. Engram can pull a public dataset from Hugging Face, search the web in bulk, have an agent research structured records with evidence, fetch the pages on a list of URLs, or have your agent write examples that must pass independent verification. It handles text, images and audio, and gives you train, validation and test files with a datasheet.

Only as accurate as you check. Engram labels with several independent passes and holds doubtful rows back, and it measures accuracy against a gold set: examples you already have, or a random sample that your own people label without seeing the AI's answer. If nothing was checked, the datasheet says so in plain words.

You do. Engram does not supply a labeling workforce. On the dataset page you create a review link and send it to teammates, domain experts or contractors. They need no account, they enter a name and label rows, and you apply their decisions when you are ready.

By default Engram keeps what it finds and records a license class per document, so raw data is not blocked. One setting keeps only open-licensed sources. The classes are a heuristic by domain, not legal advice, and what you do with the data is your decision.

Yes. Model operations take an open language model from Hugging Face, apply a known framework (bitsandbytes, llama.cpp for GGUF, llm-compressor for GPTQ and AWQ, mergekit, a LoRA distillation) and measure the result against the original on the same prompts: quality kept with an interval when you give example prompts, size, GPU memory, speed and KL divergence. You get a verdict, a served endpoint and a bundle. The model has to be public on Hugging Face (as safetensors): Engram downloads it by its id, because the work runs in the cloud and cannot see files on your machine. Private, gated and local models are not supported yet. If your model is on your own machine and can be public, push it first with the Hugging Face CLI (hf auth login, then hf upload your-name/your-model ./model-folder) and use that id. Confidential weights cannot be used yet; for those, run the same tools yourself, since the docs list the exact frameworks and settings.

Yes, as the abliterate operation, with Heretic or directional ablation. It only runs on open-weight models whose license clearly allows modification, it needs your explicit acknowledgement, it reports refusals before and after with the quality kept, and the model card and report say the model was modified. The refusal count is a keyword rule, so the report shows sample responses to let you judge it.

A classifier on a table or text is minutes. Image, audio and text-generation jobs use a GPU and take minutes to an hour on small data. Custom jobs have no fixed time cap: your agent estimates the time, the run stops at three times that estimate, and nothing runs longer than a sandbox allows.

You rent nothing and manage nothing. Engram picks the machine for the job and starts it for the run: plain CPUs for classifiers on text and tables, and a GPU for images, audio, text-to-text models and anything custom. For fine-tuning, your agent chooses the card (T4, L4, A10G, L40S, A100, H100 or H200, up to eight at once) and tells you the price first. Machines shut down when the job ends, and endpoints scale to zero when idle. Compute is ours, not yours to bring: if you want to serve a model on your own hardware, download the bundle.

Every prediction updates a tally of what the model says, and a drift check compares the last week with the mix it was tested on. When it moves, retrain on fresh data with one command. The old model keeps serving until you switch, and the run log shows whether the new score is better or worse.

No. Describe the problem in a sentence in the dashboard or the CLI, or let your coding agent do it: the CLI or the MCP server lets Claude Code, Codex or Cursor ask for a model or a dataset mid-task and wire the endpoint in. The docs cover the CLI, the REST API and the MCP tools.