Intelligent Communication Management
Enterprise communication and AI systems your organisation actually owns. Your data. Your models. Your keys. Your exit.
Since 2022
Four years building communication and AI systems.
Two products shipping
NimbleMeet and Aberon, each on a published roadmap.
Our own speech model
NimbleGPT, trained in-house for meeting acoustics.
Air-gap standard
Offline installer, offline licensing, offline updates.
The problem
Most intelligent systems run in someone else's cloud. Your meetings, your documents, your customer conversations — processed on infrastructure you do not control, by models you cannot inspect, under terms that change without you.
It works until it matters. The model is deprecated. The price moves. A regulator asks where the data went. At that moment you discover what you actually bought: access, not ownership.
The model changes
Your workflow was tuned for a version that no longer exists.
The terms change
Pricing, retention and data policy are written by someone else.
The audit arrives
You cannot prove what the system did, or who approved it.
Ownership, defined
Ownership is a property of the architecture, not of the address. It holds whether the system runs in a sealed room or in a cloud we operate for you.
01
Your data stays inside the boundary you set. No telemetry, no training on your content, no silent egress. The network path out does not exist unless you create it.
02
Model weights are delivered to you, not rented to you. Our speech model runs on your hardware. It keeps working if we disappear.
03
You hold the encryption keys. We cannot read your traffic, your recordings or your transcripts. Not as policy — as architecture.
04
Export every trace, transcript, recording and audit record in open formats, at any time. Licences verify offline. There is no server to switch off.
Deployment
Three deployment modes, one product. Move between them without migration, retraining or renegotiation.
No network path out. Offline licence verification, offline updates by tarball. Built for classified environments, banking infrastructure and isolated government networks.
Government · Defence · Central banking
Your VPC, your cloud account, your keys. Deployed by a single Ansible role or Docker Compose onto infrastructure you already pay for.
Large enterprise · Regulated industry
Single tenant, dedicated instance, your keys. We run the infrastructure; you keep the data, the weights and the right to leave with everything.
Mid-market · Distributed teams
The ladder works in both directions. Start in our cloud, move into your own network when you are ready — same product, same governance, no data left behind.
Products
Human to human. Agent to system. Human to machine. Every layer built to the same standard of ownership.
NimbleMeet
Own your meetings and their meaning. SFU efficiency for browsers, MCU compatibility for the SIP room systems you already own, and an AI scribe that writes the minutes on your own servers. Speech recognition runs on NimbleGPT, our own multilingual model.
Aberon
See every trace, stop risky actions before they execute, and prove compliance. Policy enforcement without touching agent code, PII redaction before storage, and a tamper-evident SHA-256 audit chain that admins cannot rewrite.
Voice Assistant
Built on our own speech stack, governed by Aberon, deployed inside your perimeter. No third-party speech API anywhere in the path.
Engineering notes
A short list of things we learned the expensive way. If you have deployed AI on-premise before, none of this will surprise you. If you are about to, it will save you a quarter.
01 — Speech
Speech models are benchmarked on clean, read speech. Enterprise meetings are the opposite: hard-walled rooms with speakerphone reverb, SIP legs narrowbanded to 8 kHz by G.711, packet-loss concealment inventing audio nobody spoke, two people talking over each other, and speakers switching language mid-sentence. A model that scores well on a benchmark can still be unusable on a boardroom Poly. We trained NimbleGPT on that acoustic reality, which is why one transcript can hold a room system and a mobile caller without falling apart.
02 — Media
An SFU forwards streams and stays cheap. An MCU decodes, composes and re-encodes, and its cost scales with every participant. Most platforms pick one and hand you the consequence: either the H.264 room systems you bought a decade ago become second-class citizens, or you pay for a transcoding cluster nobody budgeted. Running both paths inside one meeting, and spending MCU cycles only on the endpoints that need them, is an architectural decision made early — it cannot be retrofitted.
03 — Agents
Most observability tools tell you what an agent did after it did it. That is useful for debugging and worthless for control. Enforcement has to sit in front of the action: the tool call is intercepted, checked against policy, and allowed, blocked or escalated to a human — before the SQL executes. Everything recorded after that moment is archaeology, not safety.
04 — Audit
Whoever holds the database can edit the table, and an administrator usually holds the database. Chaining every record to the hash of the one before it changes the property: any edit breaks verification for everything that follows. That is the difference between a log you show a regulator and a log a regulator accepts.
05 — Deployment
Software that "supports on-premise" usually still reaches out on first run — for container images, model weights, or a licence check. In a sealed network none of those exist. Shipping every image and dependency inside the bundle, verifying licences from a locally signed file, and delivering updates as a tarball is unglamorous work. It is also the entire difference between a deployment that happens and one that dies in the security review.
06 — Hardware
Streaming recognition is latency-bound and runs continuously. Summarisation is throughput-bound and runs in bursts. They have different hardware profiles and no reason to share a box. A vendor who quotes one GPU figure for "the AI" has usually not run it at your scale.
Technology & security
Everything below is checkable in your own environment during a pilot. Nothing here requires you to trust us.
Install
Single Ansible role or Docker Compose. The offline bundle carries every image and dependency; updates arrive as a tarball and install through the same path.
Platform
Ubuntu 22.04 / 24.04 LTS. x86_64 primary, ARM64 available. Other distributions on request.
Media
SFU for browsers, MCU for hardware endpoints, both inside one meeting. AV1, VP9 and H.264.
Room systems
Poly, Yealink, Cisco, Grandstream and other SIP endpoints join as full participants, with per-device codec and media profiles. Cascading to external MCUs by URI or dial plan.
Speech model
NimbleGPT — multilingual, trained on speakerphone, SIP narrowband and mobile acoustics. Runs on your GPUs; no audio frame leaves the perimeter.
Hardware sizing
Streaming recognition is latency-bound, summarisation is throughput-bound. Sized separately, deployable on separate hosts.
Media security
TLS 1.3 for signalling, DTLS-SRTP for media.
Agent control
Policy evaluated before the action executes: tool restrictions, cost ceilings, human-in-the-loop approval. No changes to agent code.
Agent stacks
LangChain, AutoGen, LlamaIndex, or fully custom Python via a thin SDK.
Audit trail
SHA-256 hash chain. Each record hashes the previous one, so any edit breaks verification for everything after it. One-click integrity check, JSON export.
PII handling
Detected and redacted by Presidio before storage. Raw personal data is never written.
Identity
LDAP and Active Directory. Role-based access: administrator, owner, auditor, viewer. SAML/OIDC and SCIM on the roadmap.
Licensing
Cryptographically signed files, verified locally. No licence server, no phone-home.
Egress
Zero outbound connections in the runtime path.
Compliance
Designed to support GDPR, HIPAA and equivalent programmes, because the data does not leave your perimeter. No certification is claimed.
Who it's for
| Sector | The question they ask | Our answer |
|---|---|---|
| Government | Can it run with no path to the internet? | Full air-gap, offline licence, offline updates. |
| Defence | Who can read the traffic? | You hold the keys. We cannot. |
| Finance | Prove how the system reached that decision. | Immutable audit chain with approvals, exportable. |
| Healthcare | Where does patient data go? | Redacted before storage, never leaves your network. |
| Enterprise | We have agents in production and no single view. | One control plane: cost, policy, traces, approvals. |
Custom work
The team that ships NimbleMeet and Aberon is the team that deploys inside your network. You get the people who wrote the stack, working on your problem.
Models on your data
Fine-tuning and evaluation on your corpus, inside your perimeter. Weights are delivered to you.
Agents in production
Automation of real workflows, governed by Aberon from day one: policy, cost ceilings, human approval, audit.
Integration
Your directory, your telephony, your document systems, your existing hardware fleet.
Scope, acceptance criteria and audit requirements are agreed before a line of code is written.
Request access
We will prepare a technical briefing for your environment: deployment path, integration with what you already run, and a licensing model. Real engineers reply, usually within one business day.
Or write to info@intelcom.ai
Company
Intelligent Communication Management FZCO has built communication and AI software since 2022, from Dubai Silicon Oasis. The speech model is ours. The media stack is ours. The engineers who wrote them install them in your network.
Custom work runs on the same engineering. Fine-tuning a model on a bank's corpus and putting agents into production under policy is what our own roadmap consists of every week.
A neutral jurisdiction is part of the offer: no obligation to hand data to a third country, and no foreign parent in the supply chain.
Since 2022
Four years shipping AI products.
2 in production
NimbleMeet and Aberon, each on a published roadmap.
1 own model
NimbleGPT speech recognition, trained in-house.