What this is
An autonomous machine-learning system we built and own. You give it a natural-language goal — “build a sentiment classifier for product reviews”, “fine-tune a coding assistant”, “forecast demand for these SKUs” — and it runs the full ML lifecycle end to end: literature and dataset research, plan, experiment, evaluate, iterate, and deploy. It supports five modalities: classical ML, deep learning, computer vision, time-series, and large language models.
The system is engineered for verifiable autonomy: integrity controls catch the kind of subtle failures (data leakage, fabricated metrics, silent fallbacks, swallowed exceptions) that quietly destroy ML projects in production, and every produced artifact ships with a cryptographically-signed provenance certificate that ties it back to the run that produced it.
Why we built it
Most “AutoML” tooling either restricts what model families you can use and what data shapes it accepts, or drops the ML lifecycle into an LLM and trusts the LLM not to lie about results. Both failure modes are unacceptable for serious work — especially in regulated environments where someone will eventually have to defend the model.
We wanted to demonstrate, in our own code, that the autonomous lifecycle can be engineered for verifiability in ways that hold up to audit.
What it does
- Runs the full ML lifecycle on a natural-language goal, across classical ML, deep learning, computer vision, time-series, and large language models.
- Probes the available compute at runtime and adjusts itself to fit (CPU, GPU, memory) — no hard-coded assumptions.
- Screens inputs for personally identifiable information across every modality before training begins.
- Validates the integrity of metrics, data splits, and execution evidence around the model loop, so reported results match what actually happened.
- Resumes long runs from where they failed — sessions are crash-recoverable rather than restart-from-zero.
- Emits a signed provenance certificate for every produced artifact.
Built with
| Layer | Stack |
|---|---|
| Models | Leading frontier LLMs |
| Language | Python |
| PII screening | Industry-standard PII detection (open-source + custom rules) |
| Execution backends | Local, containerized, sandboxed, and serverless options |
| Provenance | Cryptographically signed artifact certificates |
What this means for you
If you have a portfolio of ML problems you want done well — and you cannot afford the failure modes that come with either click-to-build AutoML services or “just trust the LLM” autonomous prototypes — we can:
- Tailor this integrity-controlled autonomous-ML capability to your data, modalities, and governance posture.
- Ship the system as code your team owns, runnable on your infrastructure.
- Set up the PII, leakage, and provenance controls the way regulated work needs them.
- Train your team on the why behind each integrity control, so the system stays trustworthy as your problems and data shift.
Want to see the system run end-to-end on a synthetic problem under NDA? Contact us.