THYNKR SYSTEMS · AI RESEARCH & DEVELOPMENT

Adil Edge

Private Intelligence. Anywhere.

Lightweight text intelligence, designed to run on ordinary processors and keep suitable tasks close to the people and applications that use it.

Adil Edge is an independent research and development programme by THYNKR SYSTEMS LTD. We are exploring compact, specialised AI that can process text efficiently on CPUs, without requiring GPU or NPU acceleration, and support private offline experiences on compatible mobile devices.

The goal is not to reproduce the capabilities of the largest AI models. It is to perform selected everyday text tasks reliably, with lower resource requirements, explicit boundaries and evidence people can examine.

Status: Research & Development. No public model download, mobile app, SDK or commercial service is available yet.

Explore the research approach Discuss research collaboration

StatusResearch & Development
AvailabilityNo public model download, mobile app, SDK or commercial service yet
Current focusCapability definition, model comparison, evaluation datasets and benchmark preparation
Programme ownerTHYNKR SYSTEMS LTD

§01Why this work matters

Not every text task needs a large AI model.

Correcting a sentence, classifying a message, extracting a date or retrieving an approved answer should not automatically require a powerful cloud model.

Adil Edge investigates how much useful intelligence can be delivered through a combination of ordinary software, small classifiers and compact language models. The system is intended to select the simplest reliable approach for each supported task.

Our principle is straightforward:

Do less computation when less computation is enough.

§02Research directions

One approach, several environments.

Adil Edge Mobile — Private offline text tools

A proposed mobile edition for compatible Android devices, with iOS feasibility to be assessed separately. The aim is to process supported text tasks locally on the device's CPU after installation, without cloud inference.

Initial research priorities

Grammar and spelling correction.
Short text rewriting, shortening and simplification.
Concise writing assistance and structured text extraction.
Selected-text workflows that can be used from other applications, where the operating system permits.

We intend the strictly offline edition to avoid automatic transmission of user text, remote model fallback and content telemetry. These design objectives will be tested before any privacy guarantee is published.

Adil Edge Server — Efficient routing and knowledge assistance

A proposed lightweight server component for applications that handle frequent, bounded text requests. It may classify intent, retrieve approved FAQ content, handle straightforward conversational responses and recommend when an authorised request needs a more capable model.

A separate deterministic policy layer, not the language model, would control permissions, data handling and escalation to external providers.

Research priorities

Intent and task classification.
Deterministic and indexed FAQ retrieval.
Short, grounded responses using approved knowledge.
Policy-controlled model routing and abstention.
Measurable latency, quality and resource consumption.

Adil Edge SDK — Embeddable text intelligence

We are evaluating a future developer interface for selected text-processing operations on supported mobile, desktop and server platforms.

The intended developer experience is a small, versioned component with a clear capability contract, predictable inputs and outputs, and no requirement to send every task to an external AI provider.

SDK design, licensing and availability have not been finalised.

§03Our approach

Use the least complex method that can reliably do the job.

01 / Deterministic first
Exact matches, rules, regular expressions, templates and local text indexes can complete some tasks without invoking a model.
02 / Small classification when appropriate
Short intent and routing decisions may be better served by compact classifiers than by generating full sentences.
03 / Compact generation only when needed
A small, CPU-compatible language model can be evaluated for rewriting, correction, extraction and short answers, with bounded context and output.
04 / Policy outside the model
Permissions, privacy controls and external escalation remain enforceable in ordinary software, independently of model suggestions.
05 / Measure before making claims
Accuracy, end-to-end latency, RAM use and device compatibility must be demonstrated on defined hardware rather than inferred from model size.

§04Adil Edge Verify

AI claims should come with evidence.

We plan to publish reproducible evaluation methods so developers and users can assess what Adil Edge does well, where it fails and what it costs to run.

Our planned measurements include:

Task accuracy, failed cases and abstention behaviour.
End-to-end response latency and tokens generated per second where relevant.
Peak memory usage and CPU utilisation.
Energy use per successful standardised task, where reliable measurement is possible.
Model and runtime versions, device specifications and test conditions.
Offline operation and application-level network-isolation tests.

An Airplane Mode Challenge and open benchmark methodology are planned as public demonstrations after a working mobile prototype is validated.

Verification status: Public device benchmarks, independent privacy validation and supported-device claims are pending. No independent certification is claimed.

§05Privacy by design, with honest limits

Keep the offline experience offline.

The proposed mobile edition is designed to process supported user text on the device. We intend to avoid required cloud inference, account sign-in and automatic upload of prompts or responses.

Privacy will be assessed through application permissions, network-isolation tests and local data-handling reviews. External factors such as third-party keyboards, operating-system services, accessibility tools and device backups can affect end-to-end privacy and will need to be explained clearly.

The server edition has a different operating model: external requests may be routed only when enabled and permitted by independent policy controls.

Offline Mobile and network-capable Server are distinct products with distinct privacy boundaries.

§06Where the research stands

Research is underway. Performance is not yet established.

Defined or in preparation

Product scope and separate mobile/server capability boundaries.
Candidate model and deterministic-processing comparisons.
Synthetic evaluation datasets and benchmark methodology.
Privacy, energy and CPU-performance acceptance criteria.

Not yet demonstrated or released

A publicly downloadable or trained Adil Edge model.
Verified performance on reference phones or production server hardware.
A publicly available offline mobile application or developer SDK.
Certified privacy guarantees, measured customer savings or supported-device lists.

We expect the research scope to evolve as testing identifies which tasks can be performed reliably on modest CPUs.

§07Research questions

What are we trying to prove?

Can useful AI run without a GPU?

We are designing for CPU-native inference and will verify performance on real hardware before identifying supported devices.

Can people use it without an internet connection?

That is a key objective of the planned Mobile edition. Offline behaviour and network isolation remain to be validated.

Is Adil Edge a replacement for ChatGPT or Gemini?

No. It is focused on small, frequent and bounded text tasks, rather than broad reasoning or general-purpose cloud assistance.

Will Adil Edge be open source?

The model, software and SDK licensing approach is still under review. No public-source or redistributable-weight commitment is being made at this stage.

Is Adil Edge available to download?

Not yet. It is currently a research and development programme.

§08Research and collaboration

Interested in private, efficient text intelligence?

We welcome conversations with mobile developers, embedded-system engineers, researchers and organisations interested in CPU-efficient AI, offline privacy, reproducible benchmarks and controlled model routing.

Adil Edge is developed as an independent THYNKR SYSTEMS research programme. Potential integrations with other THYNKR products are exploratory and do not imply production deployment.

Discuss Research Collaboration

Adil Edge is an R&D programme. Features, model choice, performance, device compatibility, licensing and release availability remain subject to engineering evaluation and validation.