Safety Before Scale
Our proposed engineering approach starts with bounded execution, explicit approvals, and failure criteria before expanding capability or throughput.
About
Safety before scale.
Three voices, not one: Abbey for care, Aviva for clarity, Abi for competence. One substrate. Yours alone.
Who we are
Rooted in research. Driven by reliability.
MLAI is founder-led AI engineering work focused on assistant workflows, memory systems, and developer tools. Source references and implementation limits make the current work inspectable; proposed service engagements need project-specific acceptance criteria.
- Founder-led AI engineering
- ABI runtime source in Rust
- WDBX memory and retrieval source
- Security assertions tied to implementation scope
Values
What the company is for.
Observable Reasoning
We aim to expose provenance, retrieval context, decision checkpoints, and the operator actions that changed state. Each implementation needs its own evidence.
Performance With Proof
Our evidence policy requires workload and environment notes for latency, recall, and GPU figures, and a reproduction record before calling them measured.
Private Deployment Paths
Proposed deployment work starts by agreeing the data boundary and evaluating on-premise, VPC, hybrid, or edge requirements.
Human-Centered Control
We propose visible escalation, review, and override flows, with acceptance criteria agreed by the people responsible for the system.
Research-To-Runtime Discipline
Research translation engagements propose integration notes, constraints, and operational guidance; a prototype needs separate release acceptance.
Our mission
Make assistant workflows inspectable.
Define the source, data boundary, tool permissions, and release evidence before expanding an assistant workflow. Research and prototypes require separate implementation and operational acceptance.
- Service scope
- Proposed
- Deployment boundary
- Agreed per project
- Evidence
- Source and qualification
Principles
Four lines we do not cross.
- No autonomous write action without an observable policy boundary.
- No retrieval claim without a traceable source or confidence signal.
- No benchmark without environment notes, workload shape, and reproducibility context.
- No deployment plan that ignores rollback, incident review, and human escalation.
Thesis
Why on-device wins.
Define the data boundary
Evaluate where a workload may run, which data may leave the operator's environment, and which obligations need separate legal review.
Evaluate available hardware
Inspect the target device, runtime support, and workload before proposing local inference or acceleration.
Measure the full operating cost
Compare hardware, energy, maintenance, and provider costs using an agreed workload; local execution alone does not establish unit economics.
Entity
Registration-level facts.
- Brand
- MLAI
- Focus
- Founder-led AI engineering
- Languages
- Rust, Swift, TypeScript
- Model
- Proposed SDK licensing + integration services