Talk to Jetson — Air-Gapped AI Copilot Demo | AITI
← Back to Closed-Architecture AI Live Demo · AI Crewmate

Talk to Jetson — F-104 Virtual Copilot

A closed-architecture AI crewmate for any role in your training scenario. The F-104 is the public demo. The architecture is the point.

Press the microphone and ask Jetson anything from the F-104 reference library — procedure, aerodynamics, or air combat tactics. He listens, retrieves the relevant pages, and speaks back the grounded answer. The same software runs in three deployment modes — full cloud, hybrid, and fully air-gapped on customer GPUs inside your facility. Pick the role, provide the manuals or other authoritative data, and the system answers from them.

Why the F-104? Its flight manuals have been declassified and public for decades, the airplane is instantly recognizable across our customer base — and the founder of AITI had the personal privilege of flying it. We built this around an aircraft we love.
RAG-Grounded Voice In / Voice Out Same Stack as On-Prem Advisory Only

Same software. Different binder on the shelf.

The F-104 copilot is one example. The architecture wraps around any named position in your training environment — pick the role, provide the manuals, doctrine, procedures, SOPs, or any other authoritative data the human in that seat would be trained on. The system loads them, indexes them, and answers from them.

Air
Pilot · Copilot · Weapon Systems Officer
Armor
Tank Commander · Gunner · Loader
Air Traffic
Air Traffic Controller
Fires
Forward Observer · JTAC
ISR
Sensor Operator on an ISR Platform
Airlift
Loadmaster · Crew Chief · Flight Engineer
Naval
Bridge Crew · Combat Information Center Watch
+ Any Other
Any named position in your training environment

Offered as part of AITI's existing line of simulation systems and ground training devices, where we already serve customers operating in classified, ITAR-controlled, and air-gapped facilities.

Press the mic. Ask Jetson anything from the manual.

Hold the microphone button to talk. Release to send. Jetson answers aloud through your speakers in a few seconds. Try one of the example prompts below if you're not sure where to start.

Connecting to Jetson…

Live from a Jetson Orin Nano Super on a desk at AITI Labs in Zephyrhills, FL. Not a cloud supercomputer — off-the-shelf edge hardware under $1,000, the same class of box that ships fully air-gapped to customer facilities.

Hold to talk · release to send. Or press SPACE and hold. To end the demo, say “goodbye Jetson.”

Try one of these — procedures
— aerodynamics & air combat
You said
 
Jetson
 
What Jetson does in this demo
  • Walks step-by-step through a procedure, reading one item, waiting for the operator to call “go” or “no-go,” then moving to the next.
  • Routes emergency phrasing to the right procedure — saying “we lost the engine” flips to air restart and announces it before walking through.
  • Handles mid-sentence interruption — ask a side question in the middle of a procedure, get the answer, then resume where you left off.
  • Says “I don't have that information” when asked outside the manual. Never paraphrases from general aviation knowledge, never invents procedures, never gives clearance. Advisory only — the human in the seat makes every decision.

Same voice you just heard, now telling stories.

Two paraphrased recollections from the late Dave Groark — an F-104 pilot and friend of AITI's founder. The F-104 was America's first Mach 2 fighter. These stories sit with that reputation. Re-voiced by Jetson in the same Eryn voice from the live demo above, so you can hear how the production text-to-speech holds up in longer-form narration.

~80 seconds · Berlin Crisis

Eyeball-to-Eyeball with the East Germans

F-104s and East German MiGs playing an unauthorized game of chicken along the border. The pilots picked bad weather on purpose. One of them zoomed to seventy thousand feet — in the early Flash-Gordon-style pressure suits — got into a spin, and rode it down to thirty thousand before he caught it.

Transcript

I heard this story from my friend, the late Dave Groark — he flew the F-104 during the Berlin crisis. The squadron used to run an unauthorized game with the East German air force. They would take off from West Germany, point the airplane straight at the border, and run toward it. The East Germans would scramble MiGs to intercept. The trick was to pick bad weather — the East Germans were not strong instrument pilots. Some of them crashed trying to recover. The F-104 pilots would close on the border, turn parallel, and then zoom-climb. Back then they wore the early Flash-Gordon-style pressure suits — capstans up the side, a chin strap that pulled your head down so you could see. Marginal kit. One of Dave’s friends zoomed to seventy thousand feet, got into a spin, and rode it down to thirty thousand before he caught it. He pulled out, barely. That stunt is probably the reason the Air Force shut the whole thing down. As Dave put it, they were just a bunch of amateurs zooming airplanes up like that. But, he said, it was a lot of fun.

~3 minutes · Edwards AFB

Ninety-Two Thousand Feet

F-104 zoom profiles flown at Edwards in preparation for the lifting-body re-entry program. Push to Mach 2.26, pull, climb. Dump cabin pressure at thirty thousand. Out of afterburner at seventy. Shut the engine down at eighty. Ballistic from there — with one question on the way back down.

Transcript

Here is another one from my friend, the late Dave Groark — this time from Edwards Air Force Base. When he got to Edwards he was flying zoom profiles in the F-104 to practice re-entry — preparation for the guys who would be coming back in lifting bodies. By then they had the full pressure suits, the same kind you see in the Apollo footage, so the suit was not the limiting factor anymore. The airplane was. The F-104’s compressor inlet is magnesium. Push it too hot and you get a compressor stall — the engine chokes itself. You watched two instruments. The Compressor Inlet Temperature gauge, max around one hundred. And a slow light that blinked at you past that. Slow. Slow. Slow. The briefing at Edwards was: take it to Mach two, then zoom. Do not push past Mach two — you do not want to go too far. Dave had a friend who had already learned what too far looked like. A tremendous compressor stall. If your seatbelt was not tight, you ended up pinned to the front of the cockpit. But fighter pilots are competitive. The question on every flight was: who got the airplane the highest? So Dave pushed it. Two point two. Two point two five. Two point two six. Then, a little under Mach two point three, he pulled back and zoomed. On the way up: at thirty thousand feet, dump cabin pressure. At seventy thousand, come out of afterburner — or you will overtemp the turbine. At about eighty thousand, shut the engine down entirely. That was the only time he ever intentionally shut down a single-seat fighter’s engine. From there, ballistic. Just riding the airplane. He could see the curve of the Earth and the black of space. He topped out at ninety-two thousand feet. Stayed there four or five seconds. Then started back down — with one question on his mind. Is this engine going to relight? It did. He landed. Dave said it was a very good flight.

Original stories told by Dave Groark in his YouTube interview. Paraphrased and re-voiced as a TTS quality reference; posted with respect for his memory.

Grounded in your manual or other data — not in what the model “knows.”

A language model on its own will happily make up confident-sounding procedures. That is unacceptable for aviation. Retrieval-augmented generation fixes this by forcing the model to answer only from a specific set of documents you have authorized.

Index the documents Once

  1. Each document — flight manual, doctrine, SOP, training material, or any other authoritative data — is OCR'd into searchable text, page by page.
  2. Each document is split at three levels at once — chapter, section, and paragraph — following the document's own outline. Each level gets its own fingerprint. That way, retrieving the right chapter and finding the right paragraph are both possible from one question. Retrieving the right chapter is not the same as finding the right paragraph inside it.
  3. Each section is converted into a numerical “fingerprint” (an embedding) that captures meaning, not just words. Asking “what do I do if I lose my engine” matches a section titled FLAMEOUT — AIR START even though no words overlap.
  4. All fingerprints land in a fast vector index on disk.

Answer a question Every turn

  1. The pilot speaks. Speech-to-text turns it into a question.
  2. Before searching, the system drafts a brief plausible answer from its own training — not to use, but to look up with. A pilot's spoken question and the manual's written prose don't sound much alike; a drafted answer does. The draft becomes the search fingerprint, then it's thrown away — only the manual's actual text reaches the answer step.
  3. The question becomes the same kind of fingerprint.
  4. The index returns the handful of document sections whose fingerprints are closest — typically the top six.
  5. A reranking pass boosts sections whose headings actually contain words from the question, rescuing the right procedure when a related one would otherwise outrank it.
  6. Those exact document excerpts — with page and section numbers — are handed to the language model with strict instructions: answer only from this text; if it isn't here, say so.
  7. The grounded answer streams back to text-to-speech and is spoken to the pilot.

What this means in practice

The model is not drawing on what it “knows” about the F-104. It is reading the pages of the source documents that were just put in front of it, and speaking from those pages. Swap the documents — a different manual, a different doctrine binder, a different SOP set — and you have a crewmate for a different platform or role. Same software, different binder on the shelf.

Every spoken answer is grounded. If the source documents don't cover the question, Jetson says so and stops. He does not invent procedures, does not paraphrase from general aviation knowledge, and does not give clearance. He is advisory only. The pilot in command makes every decision.

A 30-second setup before your first turn.

Speakers usually work without fuss. Microphones often don't — this is the place most voice demos fall over. Three quick checks make this work reliably.

Step 1

Allow microphone access

When you press the mic button, your browser will ask for permission. Click Allow. If you accidentally blocked it earlier, click the lock icon next to the address bar and reset the microphone permission to Ask or Allow.

Step 2

Pick the right mic

Chrome and Edge default to whatever Windows calls the “Communications” microphone, which is often not the one you want. If Jetson doesn't hear you, open Windows Sound settings, set your headset or laptop mic as the Communications device, then refresh this page.

Step 3

Speak clearly, close in

A headset or laptop mic 12″ from your mouth works best. Open-air conference mics across the room will work but are slower and less accurate. If the status light flashes red with “mic peak zero,” the browser is connected to the wrong device — back to Step 2.

On a phone or tablet? On iOS / Safari you may need to grant microphone permission per session — tap Allow each time the prompt appears. If the mic won't engage, close any other browser tab or app that is using the microphone first, then reload this page.

About this demo — same architecture, three deployment modes.

What you hear on this page runs live from a real Jetson Orin Nano Super sitting on a desk at AITI Labs in Zephyrhills, Florida — not a cloud supercomputer. The Jetson handles audio, retrieval over the indexed F-104 manual, orchestration, and playback. Speech-to-text, the language model, and speech synthesis are reached through a small auth proxy in this hybrid mode. The system prompts, the chunked manual, the retrieval logic, and the grounded-answer behavior are byte-for-byte the same code that ships in the fully air-gapped on-premises mode.

The same software architecture ships in three deployment modes — full cloud (fastest to stand up; fine for non-sensitive training), hybrid (some components local, others cloud), and fully air-gapped on customer GPUs inside the facility. In the air-gapped mode, STT, the language model, the document index, and TTS all run locally on a workstation- or rack-mounted RTX GPU. No internet required. No data sent to third parties. The manual, the operator's voice, the questions asked, and the answers given stay on the box.

Why retrieval, not fine-tuning. A reasonable question from procurement: why not just train a model on the manual once and be done? Two reasons. First, the base model itself improves every six to twelve months — a model fine-tuned on the manual is stranded on the version it was trained on while newer base models keep getting better. Second, your manual gets revised; a fine-tuned model has to be retrained for each revision. Retrieval is the durable layer: swap the base model, swap in a revised manual, the system keeps working. The base model can change; the grounding stays.

Every retrieval is auditable — the system shows which sections and pages of the source material it pulled from, so nothing is fabricated and you can open the manual to the exact page. The path from this web demo to a hardened air-gapped deployment is a configuration change, not a rewrite.

A training capability that fits the facility.

Four concrete things this architecture unlocks for a simulator, training device, or procedure-rehearsal program — for any role you choose to wrap it around.

01

Solo trainee, full crew scenario

A trainee can run a complete crew scenario with a credible AI counterpart — copilot, gunner, ATC, observer — without scheduling a second instructor or another student in the seat next to them.

02

Procedure rehearsal at the operator's pace

Step-by-step walk-throughs with the operator calling go or no-go on each item, in their own voice. Pause, resume, and ask side questions without losing the thread.

03

Complies with how classified facilities actually work

Nothing on the network, nothing in the cloud, nothing leaving the SCIF. The manuals, the operator's voice, and every answer stay on customer-owned hardware.

04

A platform AITI can stand up around any role

The F-104 demo is one example. The architecture is general. New aircraft, new vehicle, new watch station, new SOP — same software, different binder on the shelf.

Built to fit your accreditation path.

Sovereign AI deployment — the model, the weights, the data, and the audit trail stay under your authority, inside your enclave, on your hardware. The frameworks your security officer cares about are first-class targets, not afterthoughts.

Frameworks & Accreditation
CMMC 2.0 (L2/L3) NIST SP 800-171 NIST SP 800-53 DoD IL4 / IL5 / IL6 FedRAMP High CUI handling RMF / ATO ready DISA STIG-hardened FIPS 140-3 crypto
Supply Chain & Provenance
TAA-compliant Section 889 Berry Amendment Made-in-USA AI stack No PRC-origin model weights Signed model + container provenance Model SBOM No foreign-national install access required
Networks & Edge
Air-gapped SIPRNet / JWICS / NIPRNet enclaves DDIL operations Tactical edge JADC2 / C4ISR adjacent Counter-UAS environments Cross-domain solution (CDS) compatible Data diode / one-way transfer for logs
Vendor Posture
No vendor telemetry No phone-home No license check-in Updates on signed physical media Model weights pre-staged in the enclave Zero Trust architecture
Inference Stack
Llama 3 / 3.3 Mistral Phi-4 Qwen vLLM / TGI llama.cpp / Ollama INT4 / FP8 / GGUF quantization Small Language Models (SLM) Vision-Language Models (VLM) on roadmap
Aviation & Training Doc Types
NATOPS QRH AMM / IPC / MMEL Boldface & memory items Airworthiness Directives Service bulletins AAR / hot-wash LVC / DMO CRM debrief Mission rehearsal Threat-based training Pre-mission briefing

Bottom line for your security officer: nothing leaves the enclave, nothing calls home, the model is yours, the weights are yours, the audit trail is yours, and the whole thing keeps running when the network goes down. Designed and manufactured in the United States — AS9100 + ISO 9000 + ITAR + CMMC certified, since 1997.

Want this for your platform? Read how the same architecture deploys on customer-owned Jetson, RTX, or DGX hardware inside your security boundary.
See Closed-Architecture AI →
Engage With Engineering

Simulator program. Training device. Procedure rehearsal capability.

If you are responsible for a simulator program, a training device program, or a procedure-rehearsal capability, and you'd like to discuss what this looks like wrapped around your specific platform and your specific documents, the next step is a working call with the engineering team. Bring your platform, your authoritative documents, and the security posture you operate under.

Request a Capability Briefing →
dave@aircraftinstruments.com
813-783-3361
39520 Aviation Ave, Zephyrhills, FL 33542