AvGuard is the world's first offline, air-gapped AI safety system that reads analogue cockpit instruments with computer vision, correlates live flight data against 27,684 historical NTSB accidents, and warns pilots up to 30 seconds before a situation becomes critical — without ever touching the internet.
The vast majority of General Aviation aircraft — the Cessnas, Pipers, and Beechcrafts that train every airline pilot — still rely on 1960s-era steam gauges. No trend analysis. No stall prediction. No correlation with accident history.
A modern glass-cockpit retrofit costs $50,000 – $100,000+, putting it out of reach for most owners and flight schools. And even a fully connected cockpit can't help: ChatGPT can't fly at 10,000 feet with no bars.
The result is a massive fleet — the backbone of pilot training and regional aviation — locked out of the last 60 years of safety technology.
One Pattern. Three Aircraft. Every Year.
54% of US General Aviation fatal accidents involve just three airframe families: Cessna, Piper, and Beechcraft. The same three airframes AvGuard targets on day one.
— The founding question behind AvGuard
A single self-contained unit sees your gauges, reasons about what they mean, and speaks up before the accident chain locks in — with zero modification to the airframe, zero certification burden on the owner, and zero connectivity required.
A global-shutter camera watches the six-pack at 60 fps. YOLOv11 + ResNet models digitise analogue needles at over 1,000 frames per second with >97% accuracy across 1,800+ trained gauge types.
Live telemetry is continuously vector-searched against a 151,183-chunk corpus of NTSB, ASN, and POH data. Hybrid kinetic/potential energy analysis predicts stalls and unstable approaches before they happen.
When the composite risk score crosses a pilot-tuned threshold, AvGuard speaks up with a specific, corrective action derived from the aircraft's own Pilot Operating Handbook — not generic advice.
Multi-method ensemble for gauge reading. Adaptive thresholding, colour detection, radial transforms. Cross-validated to prevent hallucination.
Llama 3.3 70B (4-bit) running locally via Ollama. mxbai-embed-large vectors in Qdrant. No data leaves the aircraft.
27,684 NTSB accident reports + 3,062 ASN global incidents + POH data, all chunked at ATA-chapter boundaries.
Total Energy Control validation rejects physically impossible CV readings. Real kinetic + potential energy monitoring in real time.
AvGuard functions in communications-denied (DDIL) environments: contested airspace, remote operations, jammed GPS. Cloud-reliant systems are non-starters here. Our target customers include regional operators, flight schools, special mission and defence.
Flight data never leaves the aircraft. This is increasingly mandatory for European and allied operators seeking to decouple critical infrastructure from foreign hyperscalers. AvGuard is GDPR-compliant by architecture, not by policy.
The AI retrieves from a bounded, versioned dataset (POH + NTSB) — never the open internet. Its reasoning is auditable, deterministic, and citable by record ID. This is the only credible path through DO-178C and FAA Class 1 EFB advisory certification.
AvGuard's technical moat is protected by two USPTO provisional patent filings, with eight additional patentable innovations identified for continuation. Each addresses a specific technical problem that has blocked certified AI from entering safety-critical aviation.
Filed with the United States Patent and Trademark Office, establishing priority date for AvGuard's two most foundational innovations. The 12-month conversion window secures a defensible technical moat for pre-seed investors.
A four-tier fallback algorithm (Exact → Fuzzy → Manufacturer → Semantic) preventing dangerous cross-correlation of safety advice between incompatible airframes. Ensures a Cessna never gets 737 advice. Critical for safety and certification.
An adaptive resource allocation method that runs 100 Hz anomaly detection on-device, only invoking the heavy LLM when composite risk exceeds 0.3. Vision outputs are cross-validated against aircraft physics limits before entering the reasoning pipeline.
Fully air-gapped LLM + Vector DB requiring zero internet. The fundamental platform innovation.
Concurrent heterogeneous queries (NTSB, ASN, ICAO) via ThreadPoolExecutor with cross-source dedup.
Real-time kinetic + potential energy monitoring. Predicts stalls before instruments register.
Vibration-resistant CV ensemble combining deep learning with classical Hough mathematics.
NTSB narratives split at sentence + ATA-chapter boundaries with aircraft metadata replication.
"Digital Ghost" plugin recreates NTSB accidents in MSFS/X-Plane via SimConnect for pilot training.
TECS + rate-of-change validation rejects implausible vision outputs before downstream use.
Unalterable flight data signing. A "Carfax for aircraft" unlocking a new resale + insurance category.
Our beachhead is pilots and owners who today have zero viable upgrade option: a $60,000 glass cockpit is out of reach, but a $2,500 clamp-on unit with a $50/month subscription is a decision a CFI can make over coffee.
Expansion markets include commercial flight training ($8B), helicopter ops, UAV monitoring, defence trainers, and MRO/insurance data licensing — a credible path from beachhead to category infrastructure.
Self-contained camera + edge processor. Clamps to the dashboard. Zero airframe modification. No avionics integration. Pilot-installable in under 10 minutes.
~65% Gross MarginContinuous database updates (new NTSB reports, new POH versions), new detection modules, flight playback, cloud logbook backup, and quarterly model improvements.
~90% Gross MarginAnonymised, aggregated flight data licensed to insurance underwriters, MROs, fleet operators, and regulators. First-ever dataset for the 80% of the fleet currently dark to digital aviation.
High-margin · RecurringOllama LLM, Qdrant vectors, multi-method anomaly detection, dashboard, WebSocket telemetry bus.
NTSB (done) · ASN (done) · ICAO ADREP + IATA integrations underway. 40% of phase complete.
Live emergency checklists pulled from aircraft-specific Pilot Operating Handbooks alongside NTSB correlation.
Jetson hardware port. 295 flight hours planned. Beta deployment to 10 US flight schools.
Advisory status certification. DO-178C readiness. Commercial launch.
// Recent Comparable
Boeing's acquisition of ForeFlight + Jeppesen (Digital Aviation portfolio) was valued at $10.55B, validating the market's appetite for comprehensive aviation data platforms. AvGuard represents the data engine for the analogue fleet — specifically the ForeFlight blind spot.
We're raising a focused pre-seed round to take AvGuard from working MVP to FAA Class 1 EFB advisory certification and first commercial deployments in US flight schools.
Target runway is 18–24 months, taking the company to paid beta across ten flight schools, the first B2B data-licensing contract, and a seed round on clear commercial traction.