Efficient on-device inference
How do we preserve useful task performance while operating within the memory, compute, thermal, and energy limits of real business devices?
LatencyPeak memoryEnergy per task
SWAPILOT RESEARCH · ACTIVE AGENDA
Intelligence that
belongs on the device.
We research how useful AI can run locally, work without continuous internet, speak the user's language, and remain under the organisation's control.
NO REQUIRED CLOUD ROUND-TRIP REPRODUCIBLE EVALUATION USER-OWNED INTELLIGENCE
01 / RESEARCH AGENDAFour questions shape
Four questions shape
the SwaPilot system.
Our research starts with operational constraints, not ideal laboratory conditions. Devices may be modest. Connectivity may disappear. Business knowledge may be sensitive. Interfaces must work for people who prefer speech over software menus.
Local business memory
How can an organisation build durable, searchable intelligence from its records without surrendering that memory to a remote platform?
Grounded retrievalPermissionsPortability
Offline workflow completion
Which business tasks can be completed end-to-end without a cloud round-trip, and where should the system deliberately ask for human review?
Task successRecoveryHuman approval
Multilingual voice action
How should a local agent understand code-switched speech, regional vocabulary, and spoken intent while turning it into structured, reviewable work?
Intent accuracyCode-switchingAction quality
02 / CURRENT STUDIES
What we are testing now.
These are active research tracks, not published performance claims. Each study moves from question to method to reproducible evidence.
Device readiness matrix
What level of useful business intelligence can different device classes sustain locally?
- METHOD
- Compare latency, memory pressure, energy use, and task quality across representative phones, laptops, and compact edge systems.
- INTENDED OUTPUT
- A transparent device-to-capability map.
Disconnected workflow completion
Can workers finish meaningful operational tasks with connectivity removed?
- METHOD
- Run repeatable retail invoice, factory inspection, service quotation, and knowledge-retrieval scenarios in offline mode.
- INTENDED OUTPUT
- Task completion and failure-recovery evidence.
Grounded local knowledge
Can the agent answer from approved business documents without inventing unsupported instructions?
- METHOD
- Use curated local manuals and policies with answerability, citation, abstention, and contradiction test sets.
- INTENDED OUTPUT
- A groundedness and safe-abstention scorecard.
Hindi–English voice workflows
Can spoken, code-switched requests become accurate business actions on-device?
- METHOD
- Evaluate intent capture, entity extraction, confirmation, and structured record creation using realistic operational prompts.
- INTENDED OUTPUT
- A multilingual task-quality evaluation set.
03 / SYSTEM HYPOTHESIS
Keep the intelligence loop local.
SwaPilot is designed around a short, inspectable path from human intent to a saved business outcome. Connectivity sits outside the core loop.
01Speak, scan, or typeHuman intent
→02On-device modelUnderstand the task
→03Local knowledgeGround the answer
→04Action enginePrepare safe work
→05Local recordReview and save
OPTIONAL CONNECTION
Updates, encrypted backup, collaboration, and integrations may connect when the user chooses. Core AI and workflows do not depend on that path.
04 / EVALUATION METHOD
Evidence before claims.
We evaluate the whole workflow, not only model output. A smaller local model is useful only when people can complete real work accurately, safely, and within device limits.
Task success
Did the user complete the intended business outcome?
First response
How quickly did useful feedback appear on the device?
Peak memory
Could the workflow run within the target hardware limit?
Energy per task
What did useful intelligence cost the device battery?
Offline completion
Which steps finished without any network dependency?
Groundedness
Was every factual answer supported by approved local knowledge?
EVIDENCE LABELS
DESIGN HYPOTHESIS
A system idea that still needs testing.
PROTOTYPE OBSERVATION
A repeatable result from the current prototype environment.
VALIDATED FINDING
A documented result from a reproducible evaluation.
05 / RESEARCH NOTESWork in progress,
Work in progress,
made legible.
We will publish concise technical notes as methods and results become reproducible. Until then, every item is clearly labelled by its current stage.
NOTE 01IN PREPARATION
Designing an AI operating layer without a required cloud round-trip
System boundaries, local-first action patterns, and where optional connectivity adds value without becoming a dependency.
NOTE 02EVALUATION DESIGN
Business memory that remains owned and portable
A practical model for local retrieval, permissions, provenance, export, deletion, and encrypted synchronization.
NOTE 03PROTOTYPE STUDY
Voice-first workflows for multilingual operational work
From code-switched speech to a structured, reviewable business action on low-cost devices.
COLLABORATE WITH USHelp test intelligence
Help test intelligence
where work happens.
We welcome device and chipset partners, research institutions, field programmes, and businesses willing to help define realistic evaluation environments.
Propose a collaboration ↗