Safe ML on sensitive customer data - without humans touching production.
FLab introduces a strict separation between human experimentation and automated production execution. Researchers iterate quickly in controlled Labs; production artifacts are produced only by audited, versioned Projects.
The problem
Many organizations want to build Data/ML products using customer data, but enabling ML work often means granting broad human access to sensitive datasets and production artifacts.
What typically happens today
- Engineers and researchers get direct access to production datasets “to move fast”.
- Training and evaluation happen in notebooks, with partial reproducibility.
- Artifacts are copied between systems with weak lineage and incomplete audit trails.
- Governance is bolted on later via ad-hoc IAM policies and process.
What the business actually needs
- Fast iteration for ML engineers and researchers.
- Strict access control, least privilege, and consistent audit logs.
- Deterministic production pipelines with versioned data + code + configs.
- Clear separation between experimentation and production outputs.
Labs (human plane)
Controlled environments for ad-hoc experimentation - designed to be productive, but intentionally not powerful enough to exfiltrate raw production data or publish final artifacts.
- Ephemeral compute with strict egress controls
- Access to curated / masked / aggregated datasets
- Experiment tracking and reproducible runs
- Outputs are proposals (code/config), not production artifacts
Projects (machine plane)
Automated pipelines that are the only path to production data access and production-grade artifacts. Everything is versioned, signed, and audited.
- Non-interactive execution (CI-style)
- Data access gated by policy + approvals
- Immutable versioned artifacts with lineage
- Promotion workflows (staging → prod) with auditability
How it works
A minimal end-to-end loop, designed to eliminate "manual production ML".
- Researchers iterate in Labs using curated or policy-approved views.
- They publish changes as code/config (e.g., Git commits), plus experiment metadata.
- Projects execute training/evaluation in automated pipelines with controlled access to raw data.
- Artifacts are produced once, versioned, and promoted through controlled stages.
- Audit logs capture who changed what, which data was accessed, and how artifacts were created.
In other words: people can explore, but only automation can produce truth.
Built for Cloud
FLab is designed as a cloud-native control plane with secure, isolated execution environments. AWS PaaS provides the primitives needed for identity, isolation, storage, and auditability.
Core building blocks
- Compute isolation for Labs and Projects (container or VM-based)
- Object storage for versioned datasets and artifacts
- Central policy + identity integration
- Immutable audit logs and observability
Note: FLab integrates with existing data lakes and orchestration tools; the key innovation is the platform model enforcing the boundary between interactive work and production execution.
Contact
Interested in hearing more, reviewing the model, or sharing your ideas (security, governance, or platform requirements)?
Email: dmytro@flab.dev