Edgent LLC · SDVOSB · CAGE 1ZSR0
A4A
Hands-free Technical Order guidance for the flight line and the hangar. A4A runs entirely on the device. It does not need a network, it does not send anything to a cloud, and every answer it gives points back to the Technical Order paragraph it came from.
Designation data
- Designation
- A4A — Aircraft Maintenance Assistance
- Function
- Assisted-reality procedural guidance for aircraft maintenance
- Reference host
- Ruggedized assisted-reality headset, Android (RealWear-class)
- Inference
- On device. No inference server, no cloud endpoint.
- Network required
- None for core function
- Source of truth
- Local Technical Order corpus, distributed as signed bundles
- Offeror
- Edgent LLC · UEI UZFLEDGS3PC5 · CAGE 1ZSR0
- Status
- Offering in development. Not a fielded system.
Note
This page describes a capability Edgent is offering to build, not work already delivered. Where it states what A4A does, read it as the design Edgent proposes and will be held to. Technical claims about third-party components are cited to their publishers. Edgent claims no past performance on this capability and none is implied.
The situation
Both hands are on the aircraft
A maintainer working a procedure has gloves on, tools in hand, and often a position they had to climb into. The Technical Order that governs the task is somewhere else — a laptop on a stand, a tablet in a bag, a printed page that will not survive the weather.
Every lookup costs a break in the task: stop, secure the tool, climb down, find the page, find the step, come back, re-establish where you were. The steps that get checked are the ones worth the trip. That is a safety and a quality problem before it is an efficiency one.
The environment does not cooperate either. Hangars and ramps are loud, and connectivity is unreliable to absent — a maintenance aid that assumes a live connection is a maintenance aid that stops working at the moment it is needed.
And the work has to be provable. Quality assurance needs evidence that the step was performed as written, tied to the airframe and the work order, not reconstructed from memory at the end of a shift.
Capability
What A4A does
Voice in, voice and display out, hands never leaving the work.
- Technical Order navigation by voice
-
Ask for a manual, a chapter, a task, or a figure and get taken there. No keyboard, no touchscreen, no removing gloves. The maintainer keeps position and keeps working.
- Step-by-step procedural guidance
-
A4A holds the place in a procedure and reads the next step on request. It tracks where the maintainer is in a long task across interruptions, so returning to the job does not mean finding the page again.
- Torque, part and specification lookup
-
Spoken lookup of the values a maintainer stops to check — torque figures, part numbers, consumables, limits and tolerances — returned with the Technical Order reference they were read from, so the value can be verified rather than trusted.
- Photo and video capture bound to the work
-
Capture by voice command, tagged at the moment of capture to tail number, work order and the procedure step in progress. The result is evidence a QA inspector can accept, produced as a by-product of doing the work rather than as a separate task.
- Remote expert “see what I see”
-
Stream the headset camera to a depot engineer or subject-matter expert who can see the same fastener the maintainer is looking at, and mark it up.
Requires connectivity
Note
Remote expert is the one function on this page that needs a network. It is designed as an addition for when connectivity exists, never as a dependency. Every other capability listed above works with the device in airplane mode.
Architecture
The model reads the manuals. It does not memorize them.
This is the decision the rest of the design follows from. A4A does not depend on a large model that has absorbed the Technical Orders during training. It runs a small model whose job is to search a local Technical Order corpus, read the passage it finds, and speak it back.
Why it is built this way
Retrieval does the work that would otherwise demand a very large model, which is what makes running on a headset realistic at all. It also produces the property that matters most in a maintenance setting: every answer is traceable to a specific Technical Order paragraph. A maintainer can check the citation, and a QA inspector can audit it.
What follows from it
Updating content becomes a data sync rather than a retraining cycle. A Technical Order revision is a new bundle, not a new model. And because the answer is quoted from the corpus rather than generated from memory, the system can tell the difference between “the manual says this” and “the manual does not cover this”.
-
The corpus is loaded onto the device
The Technical Orders for the airframes in scope are indexed and stored locally, on the headset or phone. This is the system's only source of procedural truth.
-
A spoken question becomes a search
Speech recognition runs on the device. The question is resolved against the local index to find the governing passage — the chapter, the task, the table, the figure.
-
A small model reads the passage back
The language model's job is narrow: turn a retrieved passage into a spoken answer, hold the thread of a conversation, and keep place in a procedure. It is not asked to know anything about the aircraft.
-
The answer carries its citation
The Technical Order number, revision and paragraph travel with the answer and are shown on the display. An answer without a source is not returned.
-
Revisions arrive as signed bundles
When the device next reaches a network, it takes a signed delta bundle for the Technical Orders that changed. Signature and version are verified before anything is applied, and a bad bundle rolls back rather than half-installing.
Warning
A maintenance assistant that answers confidently from a superseded Technical Order revision is a safety problem. A4A is designed to refuse: if the local corpus is known to be superseded for the task being asked about, the system says so and declines to answer rather than returning the old procedure.
The same rule applies when a question falls outside the corpus. A system that cannot answer must say it cannot — never produce a plausible one.
The runtime
A4A is built to run under WebAssembly System Interface (WASI), using WasmEdge and its WASI-NN plug-in. WasmEdge's WASI-NN plug-in provides a GGML backend that runs GGUF-format models — the llama.cpp lineage — which is what makes local inference from a portable application binary practical today.
The effect is that the application layer — retrieval, procedure state, voice interaction, capture and tagging — is one portable WebAssembly module, while the inference kernel is a native host plug-in built for each platform. One application to maintain and certify, rather than a separate native port per device.
Caution
Android-first is a deliberate choice, and it is worth stating plainly. WasmEdge supports Android on arm64-v8a in interpreter mode; it does not currently support ahead-of-time compilation on iOS, and Apple's platform rules prohibit just-in-time compilation, so a WebAssembly runtime on iPhone pays a performance penalty.
RealWear-class ruggedized headsets run Android. The operational tier therefore has no iOS constraint at all, and the restriction lands only on the consumer training tier, where inference performance matters least. The architecture is strongest exactly where the mission is.
The model
The reference candidate is Bonsai, a 500-million-parameter ternary-weight language model published by deepgrove under the Apache 2.0 licence. It adopts the Llama architecture with a Mistral tokenizer and was trained on fewer than 5 billion tokens. A permissive licence matters in a cost proposal: there is no per-seat model licensing to price.
A4A is not tied to this model. Any Apache-licensed small model that fits the device budget can take its place, and the retrieval design is what makes that substitution cheap.
Devices
Two tiers, and only one of them goes airside
A4A is device-agnostic by design. It is not device-indifferent: the operational tier and the training tier are different problems and the honest answer is different hardware.
| Tier | Reference hardware | Where it is used | Why |
|---|---|---|---|
| Operational | RealWear Navigator-class ruggedized assisted-reality headset. Vuzix-class monocular displays in the same role. | Flight line, hangar, back shop, depot. | Purpose-built for the environment: fully voice-driven with noise-cancelling microphone arrays for loud work areas, ruggedized against dust, water and drops, and rated for outdoor temperature swings. The camera is a controllable, purposeful instrument rather than an always-on sensor. |
| Training | Consumer smart glasses and standard Android phones and tablets. | Classroom, schoolhouse, simulator, familiarisation — controlled environments only. | Useful and cheap for building familiarity with the procedure and the interaction model, with none of the airside constraints in play. |
Caution
Consumer camera glasses are the wrong tool airside, and A4A does not propose them there. The objections are not stylistic:
- They are camera-forward devices in environments with camera restrictions.
- They are coupled to a consumer cloud account, which is a data-egress path that has to be argued away rather than one that does not exist.
- That coupling creates export-control exposure around imagery of military aircraft that a self-contained device simply does not have.
- They assume connectivity that hangars and ramps do not provide.
- They have no authorisation story for a DoD environment.
Proposing them for flight-line use does not survive technical evaluation. Saying so plainly is the more useful answer.
Security posture
It does not work offline. It never needs a network.
Those are different claims. “Works offline” describes a system that degrades gracefully when the connection drops. A4A has no connection to drop: inference, retrieval, speech and capture all happen on the device, and the core function never reaches for a network at all.
No egress is a stronger argument than protected egress
The usual authorisation conversation is about how data is protected in transit and at rest in someone else's environment. A4A moves that conversation: for the core capability there is no transmission, no external endpoint, and no third-party processor in the path. Imagery of an airframe stays on the headset that captured it.
That sidesteps export-control exposure rather than mitigating it, and it removes an entire class of questions from an authorisation package instead of answering them.
Updates are the only network event
Technical Order bundles and model updates arrive as signed artefacts, verified before they are applied, on the customer's own distribution timing and through the customer's own network. There is no telemetry channel and no phone-home.
Rollback is a first-class behaviour: a bundle that fails verification, or that is withdrawn after distribution, returns the device to its previous known-good corpus.
Note
Edgent is not a FedRAMP-authorized provider and does not hold a CMMC certification. Where a hosted component is required, Edgent builds within AWS GovCloud or Azure Government's FedRAMP-authorized boundary and implements its share of the controls under shared responsibility. Edgent is implementing NIST SP 800-171 Rev 2 (System Security Plan, POA&M and self-assessed SPRS score) and is completing the CMMC Level 1 self-assessment. An Authorization to Operate is granted by the government, not claimed by a vendor; A4A is architected to make that package small.
Limits
What this approach costs
A technical evaluator will probe all of these. They are stated here because pre-empting them is more useful than being caught by them.
- A small model is a small model
-
A half-billion-parameter model has limited general knowledge and limited reasoning depth. Published benchmarks for models in this class sit far below frontier systems. A4A's design answer is to not depend on that: retrieval quality, index construction and corpus coverage dominate model size for this task. Where a question genuinely requires reasoning the model cannot do, the correct output is a citation and a referral, not a synthesised answer.
- The reference model is a base model
-
Bonsai as published is not instruction-tuned, and its published implementation performs operations in 16-bit precision rather than realising the full memory saving its ternary weights allow. Instruction tuning and a quantised deployment build are development work, not properties available off the shelf. This is stated as intent.
- The device budget is real
-
Ruggedized headsets in this class carry on the order of 4 GB of RAM and 64 GB of storage, shared with everything else on the device. That sets a hard ceiling on model size and corpus size, and it is the constraint that makes the small-model-plus-retrieval design necessary rather than merely elegant. Battery and thermal behaviour across a full shift, and in a hot hangar, are engineering risks to be measured on real hardware, not assumed.
- Interpreted WebAssembly is slower than native
-
The portability of the WASI application layer is paid for in execution speed on platforms without ahead-of-time compilation. The design keeps the heavy tensor work in the native inference plug-in for exactly this reason, but the tradeoff is real and belongs in a performance test plan.
- Stale or missing content must fail loudly
-
An assistant that guesses is worse than no assistant. When the corpus does not cover a question, or the revision on the device is superseded, A4A is designed to say so and stop. Verifying that behaviour holds under adversarial questioning is an acceptance criterion, not an aspiration.
Second application
The same engine, a different user
Strip A4A back to what it actually is and it is not an aviation product. It is procedural guidance under cognitive load — keeping the place in a sequence, retrieving the right detail at the right moment, and prompting the next step without being asked twice.
In maintenance, the load is that the hands are occupied and the environment is hostile. There is a second population where the hands are free and the load is elsewhere: people living with the effects of a traumatic brain injury, which commonly affect initiation, sequencing, working memory, orientation and wayfinding, and recall of faces and names.
Those are the same functions the maintenance assistant already provides. Knowing which step comes next. Holding a place in a long procedure across an interruption. Retrieving a detail read a minute ago. Prompting the start of a task that has not been started.
Configured as assistive technology
In this configuration A4A supports task completion: a daily routine held as a checklist, a route the user has saved, a name attached to a face the user has chosen to record, a step-by-step walkthrough of a task the user has set up in advance. The content is the user's own.
The architecture carries more weight here, not less. Nothing leaves the device: no account, no subscription, no server that can see what someone needed help remembering. It keeps working with no signal — for this user that is not a convenience, it is a failure mode removed.
Note
A4A in this configuration is assistive technology for everyday task completion. It is not a medical device. It does not diagnose, treat, monitor or manage any disease or condition, it is not a substitute for care from a qualified professional, and Edgent makes no claim about health outcomes.
Edgent pursues Department of Veterans Affairs work alongside its Air Force work, so the two applications sit inside a customer base Edgent is already engaged with rather than representing a move into an unfamiliar market.
The offeror
Edgent LLC
A Texas-based modernization and AI-augmented engineering firm, SBA-certified as a Service-Disabled Veteran-Owned Small Business and eligible for SDVOSB set-aside and sole-source award.
- Legal name
- Edgent Limited Liability Company
- UEI
- UZFLEDGS3PC5
- CAGE code
- 1ZSR0
- Business type
- SBA-certified SDVOSB and VOSB · SBA VetCert 96308, valid through 2029-06-15 · Texas Veteran-Owned Business
- State of formation
- Texas
- Location
- 311 Rimrock Ct, Bastrop, TX 78602
NAICS
- 541512 Computer Systems Design Services (primary)
- 541511 Custom Computer Programming Services
- 541519 Other Computer Related Services
- 541330 Engineering Services
- 611430 Professional and Management Development Training
- 611710 Educational Support Services
- 611420 Computer Training
- 561612 Security Guards and Patrol Services
Product service codes
- D302 IT Systems Development Services
- D306 IT Systems Analysis Services
- D307 Automated Information Systems Design/Integration
- D399 Other IT and Telecom Services
- R425 Engineering and Technical Services
Relevant depth
- Enterprise architecture and platform engineering — domain-driven design, event-driven architecture, event sourcing and CQRS.
- AI/ML integration, including retrieval-augmented generation, embeddings and vector search, and LLM-powered automation pipelines.
- Federal instructional design and training delivery, with Section 508 and WCAG 2.1 AA accessibility applied throughout the learning experience.
- Legacy and mainframe modernization delivered into a customer's existing accredited cloud boundary.
Leadership
The principal is a United States Marine Corps veteran — Corporal, Military Police and Military Working Dog handler — and a service-connected disabled veteran, with more than 25 years architecting distributed systems at Fortune 500 scale and a post-graduate credential in Artificial Intelligence and Machine Learning from The University of Texas at Austin. Edgent fields an 11-person extended team across engineering, instructional design and program management.
Contact