The software our own contact centre runs on
HELION was not built to be sold first and used later. It runs InfraWeb's live queues every day — which is why it handles the unglamorous parts other platforms leave to you.
AI voice agents
Calls answered end to end, in Romanian
The part of HELION that changed our business. Voice agents that pick up, understand, act and hand over — measured on Romanian calls, not translated benchmarks.
Real Romanian, not a translation layer
Speech recognition and synthesis are chosen and tuned for Romanian, and we publish how well they actually perform rather than claiming parity with English.
Turn detection that survives a pause
Streaming endpoint detection ends the turn when the caller is genuinely done — so someone reading a meter number aloud is not cut off halfway.
Configured by operators, not engineers
Prompt, voice, opening line, transfer rules and the knowledge it can draw on are all set from the admin screens.
Escalates with context
On transfer the human agent receives the transcript, the identified customer and the ticket already opened.
Costed per call, capped per tenant
Daily spend limits apply per organisation as well as across the deployment, so one runaway campaign cannot drain the budget.
Runs on your switch
The voice path sits on FreePBX/Asterisk alongside your existing routing — the agents join queues like any other extension.
What changed
AI voice agents that actually speak Romanian
Most voice AI treats Romanian as an afterthought: trained on English, measured on English, then shipped with a Romanian label. HELION's agents are built and tested on Romanian calls, on the same switch our human agents work.
Answers the call end to end
Greets, identifies the caller from their number, searches the knowledge base, opens or updates a ticket, and closes the call — or hands over.
Escalates on rules, not on judgement
«miros de gaz», «lift blocat», «avarie» — an emergency transfers on a keyword match, never on the model deciding it sounded urgent.
Hands over with the context
The human who picks up sees the transcript, the identified customer and the ticket already opened. Nobody repeats their problem twice.
Cannot run away with your budget
Daily spend caps per organisation as well as deployment-wide, a maximum call duration per agent, and every call costed to the cent.
Starts switched off
A new agent changes nothing until you enable it. Unticked queues keep going straight to people, exactly as they do today.
Learns in a way you can read
Corrections are stored as legible lessons you can edit or switch off — not baked into a model nobody can inspect.
Caller
Bună ziua, am o problemă cu factura din luna trecută.
HELION
Bună ziua! Vă pot ajuta. Îmi spuneți codul de client?
Caller
Da, o secundă… este 4471902.
HELION
Mulțumesc. Văd factura din 14 august. Doriți detalierea consumului sau o retrimitere pe email?
Illustrative. The moment that matters is the pause after «o secundă» — the agent waits instead of interrupting, which is exactly where voice AI usually breaks in Romanian.
The caller always ends up somewhere
This is the rule the system is built around, and it is written into the code that runs the call. Every failure path — no configuration, speech recognition down, model down, budget spent, call too long, our own bug — ends in a transfer to a human queue. Never in silence, never in a dropped call. An AI that sometimes cannot help is fine. One that sometimes eats the call is not.
Modules
One platform, not a suite of integrations
Every module reads the same customer record, so context does not stop at a product boundary.
Browser softphone
WebRTC straight to the switch, with ringtones, presence, hold, transfer and recording playback. Nothing to install, nothing to update on the desktop.
Tickets & SLA
Categories and subcategories with dynamic custom fields, per-type service levels, automatic escalation and follow-up.
Email, WhatsApp & SMS
Office 365 mailboxes per department plus WhatsApp and SMS, all landing on the same conversation thread as the calls.
Customer CRM
Caller identification on the ring, full call and ticket history, deduplication on email and phone number.
Knowledge base
Internal articles with attachments and AI-assisted answer suggestions on the open ticket, so agents answer consistently.
Quality scoring
Scorecards, automatic call sampling and anomaly detection, so QA reviews a representative set rather than whatever is easy to find.
Reporting
Twenty-odd reports, because averages hide the problem
A single average handling time tells you nothing about the Tuesday afternoon your queue collapsed. These are the views our own operations team uses.
- Queue performance and abandonment
- Agent × queue matrix
- SLA by department and by agent
- Congestion heatmap by hour and weekday
- Call-by-call detail with recording playback
- First-contact resolution
- Ticket volumes by category and channel
- Agent presence and occupancy
- Abandoned calls and callback outcomes
- AI voice agent cost and containment
Security & tenancy
Built for someone else's data
HELION is multi-tenant from the schema up, because we run competing clients on the same deployment.
Tenant isolation in the data layer
Organisation scoping is enforced on the query, not in the interface, so a new screen cannot accidentally expose another tenant's records.
Five roles, least privilege
Superadmin, admin, manager, QA and agent — with manager and QA scoped to their own organisation.
Two-factor and revocable sessions
Email second factor, brute-force lockout, and sessions that can be killed centrally — a role change takes effect on the next request.
GDPR posture in writing
EU hosting, retention windows enforced by the database itself, and a controller-processor split spelled out in the contract.
Your questions answered
About the platform
What operators ask us in the first call.
It sits on top of it. HELION runs on FreePBX/Asterisk and provisions it: create a department and the inbound route, outbound caller ID and queue come with it. Your existing routing keeps working, and the AI agents join queues like any other extension.
No. The softphone is WebRTC in the browser and registers straight to the switch — the audio never passes through our servers. There is nothing to update on the desktop image, which is the failure mode that made us write our own in the first place.
Yes, that is what we do ourselves. Tenancy is enforced on the database query rather than in the interface, so a new screen cannot leak another organisation's records. An AI agent can only ever be bound to its own organisation's queues.
Every call is costed in real currency and shown per agent and per organisation. Two independent daily caps apply — one per tenant, one across the whole deployment — plus a maximum call duration per agent that bounds the worst case.
That is the point. Persona, tone, form of address, opening line, what it must never promise, what always transfers to a human, which knowledge it may use and what data to collect before handover are all set from the admin screens, in Romanian, with no engineer involved.
Every AI call is listed with its transcript, outcome, turn count and latency, and outcomes are stated plainly — resolved by the agent, transferred, the caller asked for a person, nothing was understood, budget spent. You can add your own verdict on any call, and that stays separate from the automatic score.
AI-answered. Human-guaranteed.
See it on your own numbers.
We will walk you through the platform using a queue shaped like yours — volumes, languages and all.