AVIX-AI 4-Stage Pipeline: How 19/19 HTTP 200 Works at a Vertiport

Abstract

Bird-strike risk management at K-UAM vertiports is not a signage problem or a netting problem. It is a data-pipeline problem. An eVTOL rotor operating at low RPM during ground approach is acoustically and aerodynamically distinct from a commercial turbofan, which means classical airport deterrence protocols — point-source distress calls, pyrotechnics on demand — are operationally too slow and too coarse for a vertiport pad cycling departures every four minutes.

AVIX-AI BirdThreat answers that gap with a four-stage edge-to-cloud pipeline: Identify · Classify · Respond · Monitor. The pipeline runs inference on a Jetson Orin Nano module, resolves deterrence actions at the pad level, and publishes every confirmed Animal-class entity directly into Anduril Lattice — the common operating picture layer that K-UAM traffic management will increasingly depend on through the 2027 commercial window.

At Incheon Technopark, a full validation run achieved 19/19 HTTP 200 — zero dropped API calls across a complete operational cycle (commit fbcb327, 2026-04-20). This article decomposes what that number actually means in operational terms, why the four-stage architecture matters for vertiport certification, and how the EAAF flyway geography makes a validated pipeline a non-negotiable input to Korea's K-UAM Roadmap 2030 targets.

---

1. Operational Anchor — Incheon Technopark Validation Site

The Site

Incheon Technopark occupies a coastal industrial corridor on the western edge of Incheon Metropolitan City, roughly 6 km northeast of Incheon International Airport's northern boundary. The facility hosts a mix of aerospace-adjacent tenants, a surface-level pad area suitable for eVTOL ground tests, and — critically — it sits within the foraging radius of multiple waterbird species that use the Songdo tidal flat system. That tidal flat is a designated EAAF Flyway staging site, which means bird pressure is not occasional; it is seasonal, predictable, and tracked in published EAAFP survey data.

The Technopark validation run for AVIX-AI BirdThreat (commit fbcb327, 2026-04-20) used this environment as a live operational context, not a controlled laboratory. Camera feeds, edge inference, deterrence-trigger signals, and Lattice publish calls all ran against ambient bird activity — the 19/19 HTTP 200 result was earned against real environmental noise.

Environmental Read

The Incheon coastal corridor supplies three predictable variables that any vertiport operator in the region will inherit. First, waterbird density: Black-tailed Godwits, Dunlins, Great Knots, and Eurasian Curlews stage on the Songdo mudflats during northward and southward migration windows — April through June and August through October. Second, gull pressure: Yellow-legged and Black-headed Gulls forage inland along the Incheon industrial waterfront year-round, creating a background detection baseline that a pipeline must handle without false-positive saturation. Third, raptor overflights: Eastern Buzzard and Common Kestrel appear as apex deterrence challenges because classical acoustic repellents have reduced efficacy against them.

This three-layer pressure profile — waterbird surge, gull baseline, raptor spike — is precisely the environment that stress-tests a classification pipeline's ability to discriminate and triage in real time.

Differential Factor

What separates the Incheon Technopark case from a generic Korean test scenario is the regulatory adjacency. Because the site operates within the permit envelope influenced by Incheon International Airport's wildlife hazard management program (overseen by Korea Airports Corporation), any validated pipeline must meet documentation standards that will translate directly to full vertiport certification. A 19/19 HTTP 200 result at a site with this regulatory proximity carries evidentiary weight that a rural test field cannot replicate. It is the difference between a proof-of-concept and a certification-grade provenance record.

Modern Bridge

For vertiport operators preparing for MOLIT's planned 200+ pad rollout under the K-UAM Roadmap 2030, the Technopark validation is a reference deployment they can point to in working-group submissions. The pipeline architecture, the Lattice entity schema, and the HTTP audit trail are all reproducible at new sites — the Technopark run establishes the baseline that franchise-scale deployment requires.

---

2. Problem Definition — The Four-Minute Pad Cycle and API Latency

The K-UAM Roadmap targets a 4-minute average ground cycle per pad at peak-demand vertiports by 2027. At that tempo, a bird that lands on or near the pad between a landing clearance and the next departure clearance represents an unmitigated strike risk if the detection-to-response chain has any latency gap exceeding roughly 90 seconds.

Classical airport wildlife management, as codified in ICAO Doc 9332, assumes response windows measured in minutes — a vehicle dispatch, a pyrotechnic team. That architecture is adequate for a runway with a 45-minute gap between movements. It is architecturally incompatible with a vertiport operating at high frequency.

The quantitative gap is stark. Korea's Bird Strike Information System (managed by Korea Airports Corporation) logged over 300 wildlife strike incidents at Korean civil airports in 2022 alone, the majority involving birds in the 100 g–2 kg mass range — precisely the range most dangerous to exposed rotor systems. ICAO Doc 9332 Section 4.2 establishes that detection latency, not deterrence intensity, is the primary variable in strike-probability reduction.

At the same time, the EAAF flyway runs directly over the proposed Seoul Metropolitan vertiport corridor. The flyway's Songdo–Han River axis concentrates migratory pressure over the same urban airspace where 200+ vertiports are planned. A pipeline that cannot classify and respond within a single pad-cycle window is not a wildlife hazard solution — it is a record-keeping system.

The 19/19 HTTP 200 benchmark defines what "no latency gap" looks like at the API layer. If any one of those 19 calls had returned a 408 timeout or a 503 service unavailable, the entity publish to Lattice would have been stale or absent — and the common operating picture would have shown a clean pad when the ground was not clean.

---

3. UAM KoreaTech Solution — The Four Stages Decomposed

AVIX-AI BirdThreat structures its pipeline as four operationally distinct stages, each with a defined input, processing obligation, and output handoff.

Stage 1 — Identify: A calibrated camera array captures continuous video of the pad perimeter and approach corridor. The Jetson Orin Nano edge module runs a lightweight object-detection model at the frame level, flagging candidate biological entities by motion signature and silhouette profile. This stage operates at low power draw and produces raw bounding-box candidates with no species inference.

Stage 2 — Classify: Candidates above a pixel-area threshold are passed to a secondary classifier running on the same Jetson Orin Nano hardware. The classifier assigns an Animal-class label (species-level where training data permits, family-level as fallback) and a confidence score. Classification results below a configurable operator threshold are held in a review buffer; results above threshold advance immediately to Stage 3.

Stage 3 — Respond: The deterrence-trigger logic reads the classified entity's location relative to pad geometry and active flight schedule. It selects a deterrence modality — directional acoustic, habitat-edge treatment activation, or operator alert — and executes within the same edge processing cycle. Response latency is decoupled from network availability because all deterrence logic runs on-device.

Stage 4 — Monitor: Once a deterrence action fires, the pipeline enters a post-action monitoring loop. Entity persistence, movement vector, and clearance status are tracked until the pad geometry is confirmed clear. The complete event record — entity type, location, confidence, deterrence action, clearance timestamp — is structured as a JSON payload and published as an Animal-class entity into Anduril Lattice. This is the step that produces the HTTP 200 audit trail, and it is where the 19/19 validation result was measured.

The Lattice publish is not cosmetic. It means the entity record exists on the common operating picture used by UTM nodes, tower controllers, and — in dual-use deployments — defense-adjacent airspace managers. Provenance discipline at Stage 4 is what converts a deterrence event into a certification-grade record.

---

4. Strategic Context — Why the 2027 Window Closes Fast

MOLIT's K-UAM commercial launch target is 2025–2027 depending on route class, with the metropolitan Seoul–Incheon corridor designated as the first high-density operational zone. That timeline means vertiport operators seeking pad certification in 2026–2027 are writing their wildlife hazard management plans now, not after first flight.

The KAS Part 25 compatibility requirement for vertiport ground infrastructure creates a direct regulatory hook: operators must document their wildlife hazard identification, response, and monitoring systems as part of the operational approval package. A pipeline with a validated HTTP audit trail and a structured Lattice entity schema satisfies that documentation burden in a way that ad hoc deterrence cannot.

The EAAF flyway dimension is permanent, not seasonal. The flyway's legal status under the Ramsar Convention and bilateral migratory bird treaties between Korea, Japan, Australia, and China means that vertiport operators cannot simply clear habitat without triggering environmental review. The four-stage pipeline's non-lethal, precision-deterrence architecture is specifically designed for this constraint — it displaces birds without habitat destruction, keeping the operator inside the environmental permit envelope.

The Kakao Mobility federation and the UAM Korea Travel transactional layer (App ID 6769374828) are relevant here because real-time pad status — including wildlife-clearance state — will need to propagate to the booking and dispatch layer. A pad showing "occupied / wildlife clearance in progress" is a schedulable event if the data is structured; it is an opaque delay if it is not. The four-stage pipeline's Monitor output is the upstream data source for that operational signal.

---

5. Forward Outlook

Between now and the 2027 commercial window, three milestones define the deployment roadmap for the AVIX-AI pipeline at K-UAM sites.

Q3–Q4 2026: Site-specific calibration runs at two additional vertiport candidate locations within the Seoul–Incheon corridor, using the Incheon Technopark validation protocol as the acceptance standard. Each run will target a 19/19 HTTP 200 equivalent for that site's sensor geometry and ambient bird-pressure profile.

Q1 2027: Integration of Stage 4 Monitor output with UTM-layer partners ahead of MOLIT's pre-commercial operational evaluation. The Lattice entity schema established at Technopark will serve as the reference format for wildlife-entity interoperability across the UTM ecosystem.

Q2–Q3 2027: First operational deployments at pads entering the MOLIT pre-commercial phase, with concurrent KAS Part 25 compliance documentation packages submitted to Korea Airports Corporation. The HTTP audit trail from Stage 4 will anchor the wildlife hazard management annex in each operator's approval submission.

The Acoustic Vibration Mat (90% absorption at 8–40 Hz, accelerometer audit at install) will be deployed in parallel at rooftop vertiport sites where ground-level habitat treatment is architecturally constrained — the two products operate as complementary deterrence layers across the vertiport envelope.

---

Conclusion

The 19/19 HTTP 200 result at Incheon Technopark is not a marketing number — it is a provenance record that tells vertiport operators, MOLIT working-group officials, and dual-use VCs exactly what a zero-latency wildlife detection pipeline looks like when it runs against real coastal bird pressure on a real pad. The AVIX-AI BirdThreat four-stage architecture — Identify · Classify · Respond · Monitor — was built for the four-minute pad cycle that K-UAM commercialisation demands, and its native Anduril Lattice publish means every deterrence event becomes a certification-grade entity record before the next rotor spins up. With 200+ vertiports planned along the EAAF flyway corridor and the 2027 window closing, the pipeline that already passed validation is the only one operators can point to in their approval packages today.