Technology
A practical architecture for local AI.
Small systems need thoughtful engineering.
Adil Edge is being developed around a simple idea: run the minimum intelligence needed for a task, on the hardware available, with controls that make the result dependable.
Core principles
- CPU-first execution
- Build from a reliable CPU baseline, with accelerators treated as optional optimisations.
- Offline capability
- Keep approved local functions independent of continuous connectivity.
- Deterministic where possible
- Use exact workflows for exact operations; reserve generation for tasks that benefit from it.
- Resource controls
- Set explicit boundaries for memory, context, runtime behaviour and concurrency.
- Portable design
- Share sensible core components while respecting the differences between ARM devices, desktops and servers.
- Verifiable releases
- Tie runtime, models and configurations to identifiable, reviewable artefacts.
Task execution
Understand, validate, then act.
- InterpretLightweight component
- ValidatePermitted action schema
- ActApplication-owned business logic
A request may be interpreted by a lightweight component, checked against a permitted action schema and passed to application-owned business logic. In sensitive workflows, the application — not a language model — remains the final authority.
Improving the engine
Optimisation without guesswork.
Research includes compact model selection, quantisation, memory loading, inference scheduling, workload-aware routing and optional hardware acceleration. Every optimisation must be assessed for quality, stability and resource trade-offs.
Future governed adaptation
Better over time, without uncontrolled learning.
Research
We are researching controlled improvements through configuration changes, curated retrieval and independently evaluated signed adapters. On-device training and autonomous online weight updates are not part of the initial release commitment.