Sovereign AI infrastructure
for Asia-Pacific aviation.
chien.digital is a Taiwan-based technology company. We build the first end-to-end AI stack — from radio antenna to inference server — that turns live air-traffic-control audio into structured, verifiable data for aviation safety, media, education, and airport operators.
What we build.
Aviation is digitised — except the voice between pilots and controllers. 70%+ of incidents involve a communication breakdown.
We close that gap with an end-to-end AI stack we own — antenna, receiver, model weights, GPU — all operated from Taiwan. Zero cloud in the critical path.
A 24/7 real-time ATC speech-recognition platform for Asia-Pacific airspace — live audio, bilingual transcripts, flight-track fusion, and anomaly detection.
eva one nine one heavy songshan tower runway one zero qnh one zero one eight wind one zero zero degrees one six knots cleared to land runway one zero
長榮 191 重型機，松山塔台，跑道 10，QNH 1018，風向 100 度 16 節，可落地跑道 10
dynasty seven eight two heavy taipei approach runway zero five left descend and maintain flight level one five
華航 782 重型機，台北近場，跑道 05L，下降並保持 FL150
air china one nine eight contact tower one one
中國國際 198，聯繫塔台 119.1
eva one nine one taxi via echo parking bay six behind company boeing seven eight seven
長榮 191，經 E 滑行至 6 號機位，跟隨同公司 B787
emirates three six six super taipei approach runway zero five left reduce to minimum clean speed
阿聯酋 366 超重型機，台北近場，跑道 05L，減速至最小乾淨速度
Curated samples from our 326-hour annotated dataset — the raw text is what hungASR actually produced (lowercase, run-together on purpose), and the Chinese line is hungNLP's translation from the same pipeline run.
ATC tags controller transmissions; PILOT tags aircraft readbacks. The channel column (e.g. RCTP_APP) is the frequency the message was captured on. The tag under each line is the instruction sub-type emitted by the NLP layer.
For recorded coverage and incident walkthroughs, see the YouTube channel — 686,000+ cumulative views, 5,300+ subscribers.
12 real-time detectors. Voice fused with flight track, schedule, and weather.
Each detector fuses at least three independent signals — so a single-source error can't trigger a false alarm.
Conflicting runway clearance
CRITICALController clears EVA 102 to line up and wait on 05L. Twenty seconds later, a second takeoff clearance is issued for CAL 222 on the same runway while EVA 102 has not yet departed.
- Voice — two independent clearances naming runway 05L
- ADS-B — both callsigns still on the runway surface
- Flight-phase state machine — neither aircraft has transitioned to airborne
Mayday / 7700 squawk
CRITICALPilot declares "Mayday, mayday, mayday" on approach frequency, or sets transponder code 7700.
- Voice — phonetic detector for mayday / pan-pan phrases
- ADS-B — transponder code change picked up in parallel
- Intent classifier — flagged as emergency_declaration
Altitude readback mismatch
HIGHController clears Dynasty 851 to descend and maintain flight level 120. Pilot reads back flight level 200.
- Voice — two consecutive utterances parsed as altitude_clearance + altitude_readback
- Structured parameters — 12000 ft vs 20000 ft
- Readback verification layer — mismatch
Go-around in adverse weather
HIGHAircraft on short final initiates a go-around. Controller issues a missed-approach instruction; aircraft altitude trend reverses.
- Voice — "going around" / "missed approach"
- ADS-B — altitude derivative flips from descent to climb
- METAR — crosswind component, visibility, or windshear report on the same runway
We own every layer. Not a wrapper.
Real-time speech recognition
Bilingual ATC transcripts, optimised for the Asian accents that general-purpose models miss.
Structured intent understanding
Every transmission parsed into instruction type plus structured parameters — runway, frequency, altitude, callsign.
Multi-source verification
Voice is cross-checked against flight tracks, schedules, and weather before any downstream use.
Real-time anomaly detection
Detectors for runway incursion, emergency squawks, missed approaches, read-back mismatches, and more.
Sovereign on-premise inference
Audio and intermediate data never leave our facility. Zero third-party cloud in the critical path.
Self-operated infrastructure
We run receiver stations and GPU inference ourselves, end-to-end. Bootstrapped to date.
2 stations. 1 GPU. 0 clouds.
We built the antenna-to-GPU pipeline ourselves. Audio, weights, and intermediate data never leave the facility.
Rank 13 worldwide by passenger volume (2024 · 44.92M pax).
Dedicated domestic + cross-strait hub, downtown Taipei.
Licensed spectrum receivers, line-of-sight tuned per runway, NCC compliant. Six active channels across the two stations.
Colocated on-island. Audio never leaves Taiwan, never touches a third-party cloud.
Antennas, receivers, edge servers, and colocation — all built and operated without outside capital.
8 outlets. One incident. 410K+ views.
When a China Airlines A330 aborted landing at Taipei Songshan, eight major Taiwanese outlets independently requested our tower audio. Zero outbound sales.
From 2D transcripts to a 3D airport digital twin.
Voice becomes the integration spine — the one layer existing digital-twin vendors don't have.
Stage 1 · Live real-time ATC platform
hungATC running 24/7 on 6 channels. Twelve anomaly detectors, bilingual transcripts, flight-track fusion. Media, education, and flight-safety B2B pilots underway.
Stage 2 · Regional expansion
Receiver stations in Hong Kong and Singapore, followed by Japan and Korea. Same core models, new accent fine-tuning. Incident replay as a licensed data product for news media and flight-safety investigators.
Stage 3 · Airport 3D digital twin
Upgrade the 2D transcript platform to a full airport digital twin — Google 3D Tiles terrain, real-time flight tracks, ground-vehicle positions, gate status, and weather layers. Voice-centric unified console for airports and civil-aviation authorities.
Benchmarks — Sydney Airport Maps@SYD (Bentley digital twin, saves a 200-person team 5,000+ hours/year) · Assaia TurnaroundControl (major hubs save US$70M/year in turnaround efficiency). hungATC's differentiator is voice as the integration spine — the one layer none of the existing digital-twin vendors include.
Three founders. One advisor.
Taiwan student founding team running a production system 24/7 — advised by an ISO 27001 Lead Auditor.
Serial founder. Built hungATC end-to-end — SDR hardware, model fine-tuning, 24/7 operations. Prior: founded and scaled multiple Taiwan Web3 media brands; Binance Taiwan "Best Creative" (2024).
Business operations and B2B development. Background in digital IC design and international robotics competition leadership — corporate sponsorships, team SOPs, and enterprise partnerships.
System architecture and full-stack engineering. Enterprise experience in CI/CD pipelines, technical risk review, and integration workflows.
Professor at National Kaohsiung University of Science and Technology. Former Deputy Director, AI Research Institute at III. ISO/IEC 27001:2022 Lead Auditor. Advises on AI, ATC de-identification, and infosec compliance.