-
-
CRM
-
Dashboard
-
Finance
Inspiration
It started by watching a close family member — an active MLM distributor in Mexico. She would come back from the field exhausted, with handwritten notes, unanswered WhatsApp messages, and no clear idea of how much she had actually earned that month.
The problem wasn't her effort or her talent. It was that the tools didn't exist for her. CRM and business management solutions are built for companies, not for a distributor who sells between the gym, her children's school, and her WhatsApp groups.
That led me to one question: what if there was a tool that understood her language, her context, and her daily life?
What it does
MLM Workspace AI is a conversational workspace for MLM distributors in LATAM, powered by Sherix — an AI business consultant that understands natural language in Spanish and updates the workspace in real time.
The distributor simply talks to Sherix:
"I closed a deal with Ana, $350 in protein"
And automatically:
- The sale is recorded in Finances
- Ana moves to "Client" in the CRM
- The monthly goal updates
- A visual celebration triggers
No forms. No buttons. Just conversation.
Core modules
- CRM (Clients) — Kanban pipeline with interaction history, follow-up alerts, and automatic sync with tasks
- Finances — Business and personal dashboard with configurable commission structures per MLM brand
- Content — AI-generated daily post proposals based on the distributor's profile and audience
- Practice — Objection simulator where Sherix acts as a difficult client so distributors can rehearse before going to the field
- Radar — Opportunity detector based on the distributor's daily environments and social circles
- Knowledge Base — Personalized RAG per user: PDFs, links, and notes that Sherix uses to answer product-specific questions
- Field (Campo) — WhatsApp Business integration roadmap for using the workspace directly from the field
How we built it
The entire stack runs on Google Cloud:
| Component | Technology |
|---|---|
| Primary LLM | Gemini 2.5 Flash (Vertex AI) |
| Infrastructure | Google Cloud Run |
| Authentication | Firebase Authentication |
| Database | Cloud Firestore |
| Storage | Firebase Storage |
| Frontend | React + Vite + Tailwind CSS |
| RAG | Firestore + dynamic context injected in Gemini |
| Development | Claude Code (prompt-only, no manual coding) |
Every message sent to Sherix includes:
- The last 10 messages of conversation history (cross-session memory)
- The real-time state of the workspace (CRM, finances, tasks)
- The distributor's personal knowledge base
- The psychographic profile captured during onboarding
This allows Sherix to respond with real, contextualized information — not generic answers. When a user asks "how much more do I need to sell this month?", Sherix responds with the exact number from their Firestore data.
Authentication flow: Firebase Auth with Google OAuth and email/password. Each user has a fully isolated Firestore namespace under their UID, with security rules that prevent any cross-user data access.
Gemini integration: Due to organizational Google Cloud policies that blocked traditional API keys, we implemented a Node.js proxy on Cloud Run that obtains OAuth2 tokens automatically via Application Default Credentials — without exposing credentials in the frontend.
Challenges we ran into
Google Cloud organizational policies blocked the creation of traditional API keys for Gemini. We spent significant time debugging authentication before discovering that ADC (Application Default Credentials) via a Cloud Run proxy was both the required and the architecturally superior solution.
Real-time intent detection in colloquial LATAM Spanish. "Cerré con Ana $350 de proteína", "la clienta dice que está muy caro", and "recuérdame llamar a Luis el jueves" are three completely different intents that Gemini must identify and execute simultaneously — in a single message, with partial names, slang, and missing punctuation.
JSON consistency from Gemini. When messages are complex (multiple actions in one sentence), Gemini occasionally returned malformed JSON or exceeded token limits. We solved this with a 4-layer parser: direct parse → strip markdown fences → regex extraction → graceful fallback, plus increasing maxOutputTokens to 2048 and reducing temperature to 0.4 for structured outputs.
Building as a non-technical solo founder. Every line of code in this product was generated through prompts — no manual coding. This required learning to think architecturally, write precise technical specifications, and debug by reading error messages and reasoning about system behavior.
Accomplishments that we're proud of
Shipped a complete SaaS in 90 days with 9 functional modules, multi-user auth, persistent memory, and personalized RAG — as a non-technical founder using only AI-assisted development.
Gemini understands colloquial LATAM Spanish. A distributor can say "ya hablé con la señora del gym, le interesa lo de nutrición" and Sherix correctly identifies a new prospect, extracts her context, adds her to the CRM, and responds with a follow-up recommendation.
Real workspace state injected into every Gemini call. Sherix doesn't hallucinate numbers — she reads them from Firestore in real time. When she says "you need to sell $4,130 more this month", that's the exact deficit from the user's actual data.
Commission-aware financial module. MLM distributors don't earn 100% of their sales — they earn a commission percentage that varies by brand and volume tier. Our finance module is the first we've seen that models this correctly, with configurable fixed, tiered, and point-based commission structures.
Zero manual code. The entire codebase — React frontend, Node.js proxy, Firebase integration, Gemini prompting architecture — was built through Claude Code prompts. This is itself a proof of concept for AI-native development.
What we learned
Language matters more than features. Early users didn't need more buttons — they needed to speak in their own words. "Cerré con Ana" is more natural than "register sale".
Dark mode was wrong for this audience. The first prototype had a dark theme. We replaced it entirely after realizing that MLM distributors in LATAM expect bright, familiar, high-energy interfaces — like WhatsApp, not a developer terminal.
Context is everything in LLMs. Gemini without workspace context gave generic responses. With the user's real state in the system prompt, Sherix becomes genuinely useful — a consultant who knows your business, not a chatbot.
ADC over API keys is the right architecture. What started as a forced workaround became our most robust design decision: no credentials in the frontend, no key rotation risks, automatic token refresh.
Prompt engineering is real engineering. Writing a system prompt that consistently produces valid JSON with correct action types, handles ambiguous Spanish, and maintains persona across 100+ message conversations required the same rigor as writing production code.
What's next for MLM Workspace AI
WhatsApp Business API integration — The Campo module already shows the vision: distributors should be able to register a sale, get an objection response, or generate content by simply sending a WhatsApp message to Sherix. This is our highest-priority next feature.
Imagen 3 + Veo integration — AI-generated images and short video reels for social content, directly from the distributor's product photos. The backend endpoint is already built; we're pending Vertex AI Model Garden access.
Predictive closing scores — Using the interaction history in Firestore to train a lightweight model that predicts which prospects are closest to closing, and surfaces them proactively every morning.
Multi-brand support — Many distributors work with 2-3 MLM brands simultaneously. The workspace will support multiple brand profiles with separate commission structures and product knowledge bases.
Offline-first mobile app — A React Native version optimized for low-connectivity field conditions, with background sync when internet is available.
Team features — Upline/downline visibility, shared prospect pools, and team performance dashboards for distributors who manage their own networks.
Log in or sign up for Devpost to join the conversation.