LUCA — How We Built an AI WhatsApp Assistant That Handled 344 Messages at Its First Event

1. The Problem: Communication Overhead Is Crushing Event Organizers

Picture a medical conference in Romania. Two hundred cardiologists, nurses, and researchers registered. A venue in Bucharest. Three days of sessions, workshops, accreditation credits, and certificate requirements. And a two-person organizing team fielding questions from the moment registration opens until the certificates land in people's inboxes — weeks after the event ends.

The questions are almost always the same. "Where exactly is the venue?" "Which session covers interventional cardiology?" "When does registration open on day two?" "How do I get my CME certificate?" "Is lunch included?" "Can I attend if I'm not yet a member of the society?"

Every one of these questions gets answered — by email, by phone, by WhatsApp message. The organizer has already written the answers in the event program, in the FAQ document, on the website. But participants don't read documents. They message the organizer directly, at 11pm the night before the conference, from a hotel room three cities away.

For a single event with 100 participants, this communication load can mean dozens of emails per day in the week before the event and a phone that doesn't stop ringing on event day itself. Scale that to quarterly events, multiple conferences per year, and the math becomes brutal. The organizer's expertise is in running great events — not in running a support ticket queue.

The deeper problem: this communication overhead is entirely predictable and entirely repetitive. Every question has already been answered in writing. The information exists. It just isn't reaching participants at the moment they need it, in the channel they're already using. And in Romania — as across Eastern Europe — that channel is WhatsApp.

"The organizer already wrote every answer. The problem wasn't knowledge — it was delivery. Participants needed the information at midnight, on their phones, on WhatsApp. That's exactly what AI is built to solve." — Alex, Founder, JBA Agency

This is the problem LUCA was built to solve. Not with a FAQ page. Not with a chatbot that reads from a static script. With an AI assistant that actually reads the event's own documents and responds — instantly, accurately, at any hour — from the channel participants are already in.

2. The Solution: LUCA Is Not a Chatbot. It's an Event Communication Platform.

LUCA is a full SaaS platform. The distinction matters. When most people hear "AI WhatsApp chatbot," they imagine a decision-tree bot: press 1 for venue, press 2 for schedule. LUCA is categorically different. It is a complete event communication infrastructure with an AI brain at its center.

Here is what LUCA actually does: an event organizer uploads the scientific program, the FAQ document, the accreditation details, and any other relevant files (DOCX or PDF) to their LUCA dashboard. The platform ingests those documents into a per-event vector store using OpenAI Assistants API v2 with file_search. From that moment, LUCA's AI assistant has read and understood every piece of information in those documents.

When a participant sends a WhatsApp message — to the event's dedicated LUCA number — the message hits Meta Cloud API v22.0, routes to LUCA's FastAPI backend, and the AI assistant retrieves the relevant context from the vector store and generates a grounded, accurate response. Not a hallucinated answer. Not a generic reply. An answer pulled from the organizer's own documents, returned in under 30 seconds, 24 hours a day.

Surrounding that AI core: automated WhatsApp broadcast campaigns, digital certificate generation and delivery, smart escalation for questions the AI can't answer, real-time analytics dashboards, multi-event management, and a universal registration tracker that works with any website. Sixteen frontend pages. Sixteen backend API routers. Built, deployed, and battle-tested at a real medical conference.

3. First Live Event: The Numbers

LUCA's first live deployment was at a medical conference in Romania with approximately 100 participants. Before the event, the organizing team had been handling all communication manually — email responses, phone calls, WhatsApp messages from individual contacts. We integrated LUCA before the next event cycle and measured everything.

"344 WhatsApp messages handled automatically at a single event with ~100 participants — reducing organizer emails by 84% and phone calls by 93%."

The remaining 20% of messages — those the AI couldn't fully resolve — were classified by urgency, and the organizer was notified in real-time only for those that required human judgment. Two automated WhatsApp campaigns were deployed: a pre-event reminder sequence and a post-event follow-up for certificate delivery. Both ran without any manual sending.

For an organizer who previously spent several hours per day in communication management mode leading up to and during an event, these numbers represent a fundamental change in how they work. The event ran. Participants got answers. The organizer managed the event instead of the inbox.

4. Architecture & Tech Stack: Full Breakdown

LUCA is a full-stack SaaS platform with a clear separation between a React-based organizer dashboard, a Python/FastAPI backend, and the external integrations that make the communication work. Here is the complete picture.

Frontend

The organizer-facing dashboard is built on React 19 with Vite 7 for build tooling — the fastest possible DX for a data-heavy SaaS interface. Tailwind CSS v4 handles styling at scale. Zustand manages local UI state without the boilerplate overhead of Redux, while React Query handles all server state, caching, and background refetching for live data. Recharts powers the analytics visualizations — message funnels, response time distributions, AI accuracy trends.

The dashboard spans 16 pages: event creation and management, participant list management, document upload and vector store status, campaign builder, certificate template configuration, analytics overview, cross-event comparison, escalation inbox, and super-admin views. Fully responsive — organizers manage events from their phones.

Backend

The backend is FastAPI (Python) — chosen for its async performance, automatic OpenAPI documentation, and the richness of Python's AI ecosystem. SQLAlchemy ORM with PostgreSQL handles all persistence: events, participants, messages, campaigns, certificates, audit logs. The database schema is designed for multi-tenancy from day one.

The AI core runs on OpenAI Assistants API v2 with gpt-4.1-mini as the model. Each event gets its own vector store — event documents (DOCX, PDF) are uploaded and indexed via file_search, giving the AI grounded retrieval-augmented generation over real event content. Meta Cloud API v22.0 handles all WhatsApp messaging: incoming webhooks, outgoing replies, and broadcast campaign delivery via approved message templates.

ReportLab generates PDF certificates with per-event custom templates. Resend API handles transactional email delivery — certificates, confirmations, escalation alerts. Google Calendar API enables session sync so participants can add event sessions directly to their personal calendars. The backend is containerized with Docker and deployed on Railway, with the frontend on cPanel/LiteSpeed.

"16 frontend pages, 16 backend API routers — full SaaS platform, not just a chatbot."

5. Core Features: Deep Dive

AI-Powered WhatsApp Communication

The AI communication layer is the heart of LUCA. When an organizer uploads their event documents — scientific program, FAQs, accreditation details, venue information, anything in DOCX or PDF format — LUCA ingests those files into a dedicated vector store tied to that specific event. The OpenAI Assistants API v2 file_search tool then enables semantic retrieval across those documents at query time.

When a participant messages the WhatsApp number, LUCA's backend receives the message via webhook, identifies which event the participant belongs to, loads the correct assistant with the correct vector store, and submits the query. The AI retrieves the most relevant chunks from the uploaded documents and generates a response grounded in actual event content. This is not a language model generating plausible-sounding guesses — it is retrieving from what the organizer wrote.

"AI reads event documents (DOCX/PDF) and answers from actual content — no hallucinated responses."

Per-event vector stores ensure complete isolation: the assistant answering questions about a cardiology conference cannot accidentally surface information from a corporate retreat. Each event's knowledge base is self-contained, scoped, and only accessible to that event's participants.

Automated Campaign System

LUCA includes a full scheduled broadcast system for WhatsApp campaigns. Organizers configure message templates in the dashboard — pre-event reminders, day-of logistics, post-event follow-ups — and schedule them for delivery. All templates are Meta-approved message templates, ensuring deliverability at scale.

At the first live event, two campaigns were deployed: a pre-event reminder 48 hours before the conference (venue address, parking, registration desk hours) and a post-event follow-up with certificate download instructions. Both ran on schedule without any manual intervention. Zero messages sent manually by the organizing team.

Digital Certificates

Medical conferences in Romania require CME (continuing medical education) credit certificates. Historically, these were generated manually, printed, or emailed one-by-one after the event — adding hours of administrative work to an already exhausted post-event period.

LUCA automates the entire certificate pipeline. ReportLab generates PDF certificates from per-event custom templates — logo, event name, participant name, CME credits, dates. After the event concludes and attendance is confirmed, certificates are delivered automatically to each participant's email address via Resend API. Each certificate is tied to the participant's attendance record in the database. The organizer approves and triggers the batch — that's it.

Smart Escalation

Not every question has an answer in the documents. A participant might message about a genuine emergency, a complaint requiring human judgment, or a question that falls completely outside the event's scope. LUCA handles these cases with a smart escalation layer.

When the AI determines it cannot confidently answer a question, LUCA classifies the message by urgency and performs sentiment analysis. Critical messages — those indicating distress, urgent logistics issues, or strong negative sentiment — trigger an immediate real-time notification to the organizer. Standard unanswerable questions are queued in the escalation inbox in the dashboard for review during normal working hours.

"Smart escalation: questions the AI can't answer are classified by urgency and routed to the organizer in real-time."

This means organizers are never completely out of the loop — they just only hear about the things that genuinely require them. The 80% that the AI handles fully never reaches their attention at all.

Real-Time Analytics Dashboard

Every message, response, campaign send, and escalation event is logged and surfaced in the organizer's analytics dashboard. The Recharts-powered visualizations include: message volume over time, AI resolution rate vs. escalation rate, average response time trends, campaign open and engagement metrics, and cross-event comparison for multi-conference organizers.

The analytics layer serves two purposes: operational visibility during the event (is the AI performing well? are there unusual spikes in questions about a specific topic?) and strategic insight across events (which question categories repeat? what should be added to the FAQ document for next time?). The dashboard turns communication data into organizational intelligence.

Multi-Event Management

LUCA is not a per-event tool — it is a platform. A single organizer account manages multiple events, each with its own participant list, document store, campaign schedule, and analytics. A single shared WhatsApp number handles routing across all events: when a participant messages in, LUCA identifies which event they're registered for and routes accordingly.

"Single shared WhatsApp number serves multiple events with intelligent routing."

Cross-event participant profiles add a layer of personalization: returning participants — those who attended a previous event — are recognized. LUCA can greet them by name, reference their history, and deliver a materially different experience than a first-time attendee receives. For organizations that run recurring event series, this creates continuity that feels like a relationship, not a transaction.

Universal Registration Tracker

Event organizers use every conceivable tool for registration — custom websites, Google Forms, Eventbrite, proprietary medical society platforms. LUCA cannot dictate which registration system an organizer uses. So instead of requiring migration, LUCA provides a Universal Tracker: a JavaScript snippet that organizers embed in any registration page, regardless of platform.

When a participant completes registration on the organizer's existing system, the tracker fires, captures the registration event, and immediately sends a WhatsApp confirmation to the participant: "Your registration for [Event Name] is confirmed. Save this number — your event assistant LUCA will be in touch." From that moment, the participant is enrolled in the LUCA ecosystem for that event.

6. Privacy & Smart Architecture Decisions

Building on the WhatsApp Business API with AI-powered communication requires deliberate architectural decisions around data isolation, reliability, and compliance. Several of these decisions shaped LUCA's architecture from the ground up.

• Per-event vector stores: Each event's documents are stored in a fully isolated OpenAI vector store. There is no shared retrieval pool where information from one event could appear in responses for another. This is a hard architectural guarantee, not a soft filtering rule.
• Template-based messaging with auto-recovery: WhatsApp Business API requires pre-approved message templates for broadcast campaigns. LUCA stores template IDs and handles stale reference scenarios automatically — when a template reference expires or changes, the system recovers without manual intervention.
• Returning participant detection: Participant phone numbers are normalized and matched across events. The system identifies returning participants at message receipt time, before the AI processes the query, enabling personalized context to be injected into the conversation.
• Audit logging for super-admin: Every significant system action — document uploads, vector store creations, campaign sends, escalations, certificate deliveries — is logged with timestamp, actor, and event context. Super-admin views give complete operational visibility across all organizer accounts.
• Romanian-first: The entire platform UI and all participant-facing communication are in Romanian. This is not an afterthought — the AI assistant's system prompts, the template messages, the certificate templates, the escalation notifications — everything is designed for the Romanian event market. Localization is not a translation layer; it is baked into the product DNA.

7. The Development Journey: v1.0 to v9.0

LUCA was not designed in full and then built. It was iterated into existence over six months of continuous development, with real event deployments driving every major architectural decision. This is what AI-first development actually looks like in practice — not waterfall planning with a requirements document, but rapid shipping with real users providing the feedback loop.

Each version increment was driven by a real question: what broke in production? What did an organizer ask for after their first event? What edge case in the WhatsApp Business API required a workaround? The roadmap was written by reality, not by a product planning session.

The AI tools available during this development period — GitHub Copilot, Claude, and others — materially accelerated the iteration velocity. Boilerplate code that would have taken days was generated in minutes. Complex integration patterns for the Meta Cloud API were reasoned through with AI assistance. This is not a footnote; it is a core part of what made a six-month sprint from v1.0 to a production-hardened v9.0 possible with a small team.

8. Lessons Learned: What Building Vertical AI Actually Teaches You

9. Why This Matters: AI Will Transform Every Industry's Communication

LUCA is a medical conference assistant. But the pattern it demonstrates is universal. Every industry has a version of the event organizer problem: a small team, a large number of stakeholders, a predictable set of questions, an existing body of written knowledge, and a communication channel that stakeholders already use. In most industries, those questions are still being answered manually, one by one, by humans who should be doing higher-value work.

The legal firm answering the same onboarding questions from every new client. The real estate agency fielding the same property questions from every lead. The hospital answering the same pre-appointment questions from every patient. The municipality responding to the same citizen requests every day. Every one of these is a LUCA-pattern problem: structured knowledge, predictable queries, high-frequency low-value communication that consumes human attention at scale.

What LUCA demonstrates is that the technical infrastructure for solving these problems exists today, is accessible at reasonable cost, and can be deployed in a specific vertical with results that are measurable from the first event. 344 messages. 84% fewer emails. 93% fewer calls. These are not projections — they are measured outcomes from a real deployment.

The question for any business is not whether AI can handle their communication overhead — it demonstrably can. The question is who will build the vertical-specific system that integrates AI into their specific workflow, their specific data, their specific communication channels. That is precisely what a Fractional AI Officer from JBA Agency identifies and delivers: not generic AI adoption advice, but specific, measurable AI infrastructure built for your operations.

"LUCA is not a product story. It is a proof of what's possible when AI is integrated into a specific workflow with real data, real constraints, and real measurement. Every industry has its version of this problem waiting to be solved." — Alex, Founder, JBA Agency

Event management was the vertical we started with because the pain was vivid, the constraints were clear, and the measurement was straightforward. But the architecture — AI reading your documents, responding in your stakeholders' preferred channel, escalating intelligently, automating repetitive tasks, giving you analytics — translates directly to adjacent verticals. Healthcare communication. Legal client management. Corporate onboarding. Property management. Education.

If you are interested in exploring what this looks like for your business, the conversation starts at /fractional-ai.html. If you are an event organizer in Romania and want LUCA specifically, visit getluca.ro or reach out directly via jbagency.ro.