AslasChatAI-powered chatbot SaaS — automated customer interactions and NLP-based lead capture.

The challenge

Small businesses wanted to deploy AI chatbots on their websites without engineering teams — and they needed those bots to actually capture qualified leads, not just answer FAQs. Existing tools were either too generic, too expensive, or required code.

What I built

Built a multi-tenant SaaS platform on NestJS + Next.js: scalable REST APIs, Firebase authentication, secure token handling, and role-based access. Integrated Google Gemini for NLP-based extraction of names, emails, and phone numbers directly from natural chat flow. Real-time chat surfaces, dashboards, and analytics built with MongoDB and Tailwind.

• Multi-tenant SaaS architecture with isolated workspaces
• Firebase auth + RBAC + secure token handling
• Gemini-powered NLP extraction for names / emails / phones
• Real-time chat, conversation dashboards, and analytics

Key technical decisions

The choices behind the choices — and what tradeoff each one made.

• Firebase Auth over rolling my own

  Auth is one place where boring infrastructure wins. Firebase covers email, social, and OTP out of the box — the time saved went into the chat and NLP pieces that actually differentiate the product.

• MongoDB for chat documents, indexed aggregations for analytics

  Conversations are naturally document-shaped. Mongo's aggregation pipeline handles the per-tenant analytics rollups without a separate warehouse — fast enough at the scale this product targets.

• Gemini for lead extraction, not generation

  Using the LLM for structured NLP (entity extraction) instead of free-form output keeps results deterministic and cheap. A simple Zod schema validates every extraction before it lands in a tenant's dashboard.

Outcome

Tenants spin up bots in minutes, see conversations in real time, and watch qualified leads land in their dashboard automatically. The platform was the foundation of my Final Year Project at FAST University, completed in 2026.

Lessons & what I’d do differently

• Build tenant isolation into the data model on day one — adding it after first paying customer is the kind of refactor that delays everything else.
• Real-time UI feels essential, but a thoughtful 'last 24h' view covers 80% of what tenants actually check. Don't over-invest in WebSockets early.
• NLP extraction quality lives or dies by the prompt + schema pair — treat it as a versioned artifact, not a string in the code.