Recepción24

The challenge

Most small businesses lose customers for a boring reason: they don't pick up the phone. The call comes in while the owner is serving someone else, driving, or closed; the customer hangs up and calls the next name on the list. Same with the WhatsApp messages that sit on "seen". A human receptionist is expensive and only covers one shift; an answering machine resolves nothing.

The challenge for Recepción24 was building an AI receptionist that actually resolves the call — understands the customer, answers with the business's own information, takes a booking or a message and alerts the owner — with latency low enough to feel natural, and with the same "employee" answering on voice and WhatsApp with the same persona and memory. All of it as a multi-tenant platform: onboarding a new business is configuration, not code.

What we built

• A text-in / text-out "brain" on LiveKit Agents (Python): streaming transcription, language model and speech synthesis chained for a fluid phone conversation, with end-of-turn detection and barge-in.
• Employees as configuration, not code: each AI employee's persona, business knowledge, tools and rules live in data (Azure SQL), not the binary — onboarding a new client means seeding rows, not deploying.
• Swappable channels: because the brain is text-in/text-out, voice (SIP via Telnyx and Twilio) and WhatsApp (Meta Cloud API direct) are adapters that share the same persona, knowledge and conversation memory.
• Business tools: capture a lead, notify the owner by email, and hand off to a human — the agent doesn't just chat, it executes actions and leaves a record.
• A .NET / Blazor control plane: managing employees, prompt versions, knowledge, number bindings, and reviewing conversations, leads and recordings — separate from the brain so content can iterate without touching the runtime.
• Call recording and transcription with a recording disclosure, replayable from the panel, for quality control and traceability.
• Data in the EU (Azure West Europe) and a GDPR-first design — a selling point, not a footnote.

Technical decisions that mattered

1. The brain is text-in / text-out

The architectural decision that holds everything up: the core reasons over text and nothing else. Voice and WhatsApp are adapters around it. That means adding a new channel (or switching voice provider) doesn't touch the agent logic, and the same employee can answer a call and then continue the conversation over WhatsApp without losing the thread.

2. Employees as configuration

A new business isn't a code branch: it's a handful of rows. Persona, knowledge, tools and guardrails are resolved at runtime from the number the call comes in on. That lets the platform scale to many clients without every onboarding being a deploy, and the content is edited by an operator from the panel.

3. Voice over SIP with two carriers

Numbers come in as SIP trunks from Telnyx and Twilio into LiveKit, giving numbering flexibility (including Spain) without locking into a single carrier. WhatsApp goes through Meta's Cloud API directly, no middleman, for first-class messaging.

4. Latency treated as a product

On a call, silence kills. The pipeline was measured turn by turn (end of speech, first model token, first synthesis audio) and the work went into the real bottlenecks — because the difference between "receptionist" and "robot" is decided in seconds.

Outcome

Recepción24 is in production answering real calls: its first "employee" is Reparaman, answering on voice and WhatsApp; ListingOK is the second. The same runtime serves different brands by changing only the configuration — which is exactly the platform's promise.

The most underrated decision in hindsight: separating the brain from the channels. Treating voice and WhatsApp as adapters of a text-in/text-out core is what turns "an AI receptionist" into "a platform of AI receptionists" — and what lets you add the next channel or the next client without rewriting anything.