Trading floors, airport operations centres, control rooms and courtrooms all run on voice. Aquenta builds the platform those rooms talk on, and the sovereign AI that listens to them. Every line can be recorded, and every output traces back to its source.
Our customers run trading floors, airports, airspace and courtrooms. They don't need technology that is clever in a demo. They need systems that are right, that say so when something is wrong, and that leave a trail anyone can follow.
Sovereign by default
Runs on your hardware, in your data centre, air-gapped if you need it. No audio and no inference leaves your walls.
Provenance on everything
Call lines stream to a compliance recorder. Every event, alert and translated line carries its source, its model version and the checks it passed.
Humans hold authority
AI interprets. Deterministic rules and people decide. Our systems advise and record; they never take control.
Honest when it matters
When the signal is unclear we say so. When recording stops, the turret says so, in a banner nobody can dismiss.
Concord
One turret platform for every room that cannot go quiet.
Many organisations. One platform. No crossed wires.
Turrets grew up on trading floors, but the same needs exist wherever people coordinate by voice under pressure: instant lines, open channels, push to talk and a record of what was said. Concord puts trading floors, airport operations centres and security command rooms on one software platform. Each organisation gets its own turret, its own people and its own look. When they need each other, it is one press away.
01 / 05 · One platform
Every organisation, one turret system
The sign-in screen lists all three companies. One Concord server hosts them all, and each signs in to its own space with its own people, lines, branding and rules.
Multi-tenant by design
Each organisation is a world of its own, until it chooses otherwise.
Separate by default. Every company has its own users, lines, numbers, layouts and audit trail. Its administrators never see another company's people.
Its own identity. Name, logo, colours, console type, ringers and national call tones are set per company. Even the admin console wears the company's colours.
Consent on both sides. Two companies can find and call each other only when both allow it. Untick one box and the link is gone.
One numbering plan. Every company keeps its short internal numbers. Across the platform, 88 plus the company code plus the number reaches any desk that allows it.
One operator view. A separate platform console, behind two-factor sign-in, adds companies, assigns numbers and watches capacity and health across the deployment.
On a single media node
~3,000concurrently active audio streams
2,000listeners on one hoot, clean, 8.2 ms tick p99
1,000simultaneous two-party G.711 calls
0.07 µsadded per packet for SRTP encryption
Measured on one 12-core server (Apple M4 Pro), load generator on the same machine. 3,100 streams across 100 hoots used about 3.4 cores, under half the mixer's limit. Need more? Conferences span nodes, and nodes start and stop as load moves.
Architecture
Rust where audio lives. C++ where people touch it.
Signalling, media and the console run in separate planes. Audio never waits on a garbage collector.
Turret consoleC++20 and Qt 6, touch first. Linux, Raspberry Pi kiosk and macOS
Media nodeRust real-time mixer, SRTP, Opus, G.722 and G.711, one encode per conference
SIP gatewayRust, UDP, TCP and TLS. Desk phones join the same lines and hoots
Control planePython and FastAPI: tenants, entitlements, floor control, routing, licensing
Tests2,300+ control plane, ~300 client, 350+ Rust, plus real calls end to end
Security
Encrypted by default, from sign-in to speaker.
Every session gets new keys, and every admin action lands in the company's audit log.
AudioSRTP with AES-GCM, checked against RFC test vectors and libsrtp
SignallingHTTPS and secure WebSocket for turrets, SIP over TLS for phones
Peoplescrypt passwords, turret passcodes that lock after five tries, 2FA for operators
SecretsRecorder keys and carrier credentials sealed at rest under a separate key
SoftwareEd25519-signed releases and licences, staged rollout and rollback
Recording
Built to be recorded, and to say so when it is not.
Each seat streams its lines to a compliance recorder, with an optional second recorder at a DR site, over mutual TLS 1.3 with a certificate per seat. Recording is set per company.
REC OKEvery recorded line on this turret is reaching the recorder.
DEGRADEDOne recorder or stream has a problem. Recording continues on the rest, and the operator can see it.
NOT RECORDINGA banner the operator cannot dismiss.
The recorder integration is in acceptance testing with our recording partner.
Deployment
From box to working desk in an afternoon.
One signed installer brings up the control plane, media node and SIP gateway as self-restarting services.
HardwareLinux touch panels, Raspberry Pi kiosks and Mac desks run the same console
PairingA new turret joins its company with a six-digit code from the admin console
ScaleAdd media nodes for capacity. Kubernetes manifests and autoscaling included
LicensingBy companies, users, seats, channels, phones and nodes, with live usage shown
Coming next, not in the product today: SIP trunks to carriers and counterparties, a replicated control plane for high availability, Android and iPad consoles, physical key panels, and a gateway to real radio networks.
Screens are real Concord software running on our demo platform. Aquenta Capital, Northgate International Airport and Hillcrest Estate are demonstration companies with invented people and numbers.
Verbasignal
Every word on the radio, turned into an event you can trust.
Signals from the verbose.
Ground operations run on voice. Pushbacks, start-ups, taxi clearances and line-ups are spoken, heard and then keyed in by hand, late or not at all. Verbasignal sits beside your existing radio and turret systems, ours or anyone else's, listens to every channel, and emits a clean, canonical event stream your operational systems can consume directly.
Canonical events, not transcripts. Typed movement milestones with callsign, stand, position, timing, confidence and full provenance, published against a versioned schema.
Episodic and semantic memory. Built around a memory of what your site has heard and confirmed (every operator correction becomes an episode) and what it knows (airline aliases, stands, taxiways). It gets sharper with every shift.
AI interprets, rules authorise. A deterministic state machine is the final authority. A model can never write operational state on its own.
Emit early, revise honestly. Events go out fast as provisional, then settle. Unclear calls are delivered as unresolved, never silently dropped or guessed.
Built for real radio. Push-to-talk gating, cross-position de-duplication and defences against silence hallucinations, before a single word is transcribed.
Speaks back, in your passengers' languages. Events become announcements written and voiced on-prem, with every flight, gate and belt number checked before anything is played.
Verbasignal · Ground positionsLive
Transcript
Canonical events
Downstream
Every event can speak. In 23 languages.
The same event that reaches your operational systems can become a passenger announcement. An on-prem language model writes it from the event and the flight board, a deterministic check makes sure no flight, gate or belt number changed on the way, and it is translated and voiced for the PA, the display boards and the app.
Verbasignal · AnnouncementsLive
1 Event
Flight board · from MumbaiBaggage · belt 5
2 Written on-prem
flight 417 keptbelt 5 keptcategory on list
3 Translated & spoken
PA voiceDisplay boardsApp & SMS
Skyweave
One sky. One picture. Ask it anything.
Traffic, weather, NOTAMs and infrastructure, woven into a single operating picture.
Air navigation and airport teams juggle separate systems for surveillance, weather, notices and capacity, often split across several flight regions. Skyweave fuses them into one read-only picture for every role, and puts an AI copilot on top that you can simply talk to.
Talk to the airspace. A voice and text copilot that answers from live data, flies the map to what matters and drafts situation briefings in seconds.
Grounded, never invented. The copilot can only speak about flights, stations and hazards that exist in the data. Every answer cites its source.
Correlation and prediction. Aircraft inside hazard cells, flight-category changes at the field, and delay risk per airport with a stated reason and a suggested mitigation.
Read-only by design. Skyweave sits above your air traffic systems with no write path into them. It informs; it never issues a clearance.
Replay anything. Live, simulated and recorded feeds share one schema. Scrub the timeline to see the picture as it stood at any moment.
Skyweave · Operating pictureLive
TrafficIn hazardConvective cellAerodrome
ADVISORY
Interpres
Every voice heard. Every word on the record.
Live transcription and translation, built for courtrooms and hearings.
When a witness speaks a different language from the court, justice slows down and the record gets harder to check. Interpres streams faithful, court-appropriate translation onto the bench screen as people speak, and keeps a source transcript beside every translated line so nothing has to be taken on trust.
Live, in about three seconds. Provisional text appears as words are spoken, then settles into final text at each pause.
Transcribe first, then translate. The source-language transcript is kept and checkable on its own, so every translated line traces back to what was actually said.
Legal register, protected details. A refinement pass applies legal terminology and guards the things courts care about most: names, numbers, dates and negation.
Case memory, strictly walled. Later hearings of the same case remember its people, places and terms. Context never overrides the spoken word, and nothing crosses between cases.
Sovereign and tamper-evident. Fully on-premise with no internet dependency, active-active resilience and a signed audit envelope for every session.
Interpres · Courtroom 4 · Session in progressLive
Source: PortugueseTarget: English (legal)Speaker: WitnessCase glossary loaded
Source transcriptlistening
Translationwaiting
speech → screen
0.0 s
audit · audio ref · source transcript · translation · timestamps · confidence · model + glossary version
300,000+real ground-radio transmissions processed on site
~3,000live audio streams mixed on a single Concord media node
~3 starget from spoken phrase to translated text
0bytes of inference leave your infrastructure
In the field
Built on real operations, not lab data.
We work with a major international hub airport, a national air navigation service provider and a national judiciary. We can't name them, but we can show you what we built and how it runs.
On siteMajor international hub airport
Verbasignal on the apron
Installed beside the airport's existing turret and voice-recorder estate, reading ground-control radio across every position, around the clock.
Corpus3 weeks of continuous ground radio, 300k+ transmissions after de-duplication
CaptureLive turret RTP and per-transmission recorder audio, normalised in one ingest path
Pre-ASR filtering~23% of audio gated as non-speech, ~42% removed as cross-position duplicates
EventsICAO-aligned movement milestones, resolved against the live flight board
DeploymentAir-gapped GPU appliance, offline-signed licence, no transcript text in logs
DemonstratedNational air navigation service provider
Skyweave across a national airspace
One operating picture spanning several flight information regions, with a multilingual copilot for operations staff and a role-based view for each stakeholder.
SurveillanceADS-B plus ASTERIX radar and fused-track feeds, from multicast or recorded capture
WeatherMETAR, SIGMET and AIRMET, mapped to hazard polygons and flight categories
ThroughputA full day of national surveillance ingested in under 30 minutes
ReplayAny moment reconstructable, at up to 3600× real time
UpdatesSub-2-second push to every console
SafetyNo write path into ATC systems, audit-logged AI outputs, human sign-off
ArchitectedNational judiciary
Interpres for a nationwide court estate
Sized for dozens of courtrooms in session at once, translating witness testimony from many languages into the court's language and legal register.
LatencyP95 ≤ 3 s, P99 ≤ 4 s from end of phrase to bench screen, tracked per stage
PipelineStreaming ASR, draft translation, then a constrained legal-refinement pass
QualityPer-language scorecards on human-verified test sets, gated canary releases
RecordAudit envelope per utterance: audio ref, transcript, translation, versions, corrections
ResilienceSingle-site HA or two active-active sites, sessions reconnect on failover
SovereigntyFully on-premise, mTLS, signed session requests, zero internet dependency
How we work
Measured before it ships. Accountable after.
We don't sell a model. Models change every quarter. What lasts is the data, the evaluation and the guard rails around them, so that is what we build first.
Listen to your reality
We start with your real recordings, feeds and documents, not a vendor benchmark. Your site's vocabulary and edge cases shape the system from day one.
Build the gold set
Human-verified reference data for every language, channel and event type. It becomes the yardstick every release is held to.
Gate every release
Shadow runs, canary rollouts and per-language scorecards. Models are swapped only when they win on your data, never because they look good in a demo.
Run it where you are
Deployed on your hardware with offline licensing, deterministic replay and a full audit trail. You own the record.
Get in touch
Let's hear what your operation sounds like.
Tell us about the rooms, voices and records you depend on. We'll show you a turret built for them, and what their traffic looks like as structured, trustworthy signal.
We reply within two working days. Briefings can be held under NDA.