Brainfish : Scalable AI Chatbot for Domain-Specific, High-Speed Customer Support

General AI SaaS Platforms

Brainfish – Domain-Specific AI Chatbot for Fast Self-Service

Brainfish needed a scalable AI chatbot solution to serve platforms like Airtasker and Mad Paws—and we delivered. We built an intelligent assistant trained on complex FAQs, help articles, and live chat data that can instantly answer customer queries across domains. It uses semantic search and context retention to surface precise answers, even across different products or topics. The bot handles over 85% of common queries autonomously, reducing support costs and response times. With multilingual capabilities and self-learning feedback loops, Brainfish delivers lightning-fast support with minimal human intervention.

Project Overview

Client: Brainfish (Customer self-service AI platform used by companies like Airtasker & Mad Paws)
Challenge: Manual query handling couldn’t scale across different knowledge domains
Goal: Build a unified AI chatbot to:
- Instantly answer high-volume FAQs across multiple domains
- Use semantic understanding to serve accurate responses, even to complex questions
- Reduce support tickets while improving user experience and deflection rates
Team: 7 (2 NLP Engineers, 2 Backend Devs, 1 UX Lead, 1 QA, 1 PM)
Timeline: 4 months (Training Phase → Beta Rollout → Full Integration)

“GenX built us a support engine that thinks, adapts, and scales. Brainfish now delivers expert answers—faster than any human ever could.”

Co-Founder & CTO, Brainfish

The Challenge

Critical Pain Points:

User questions spanned multiple product areas, requiring intelligent context switching
Legacy bots relied on keyword triggers, often surfacing irrelevant or incomplete answers
High support team load due to repeat queries and lack of real-time automation

Technical Hurdles:

Creating domain-specific embeddings that could generalize across help content
Handling layered questions that required memory of prior turns in conversation
Ensuring quality answers in different languages with minimal retraining

Tech Stack

Component	Technologies
Search & Retrieval Models	FAISS, OpenAI Embeddings, BM25, Pinecone
Dialogue & Response Models	GPT-4, LangChain, Rasa, Custom Prompt Templates
Data Infrastructure	Firebase, PostgreSQL, Redis
Frontend SDKs & Chat UI	React, Web Components, Flutter SDK
Monitoring & Analytics	Mixpanel, Sentry, Real-time Feedback Tracker
Language Support	Google Translate API, DeepL, spaCy

Key Innovations

The chatbot learned from help docs and live chat logs to resolve queries instantly. Semantic search provided accurate answers across products. With multilingual and self-learning features, Brainfish scaled support with minimal staff.

Cross-Domain Semantic Matching

Delivered accurate answers even when user phrasing varied significantly

Result: 87% query resolution rate without human intervention

Memory-Driven Conversational Intelligence

Tracked intent across multi-turn dialogues and adjusted answers dynamically

Result: 3.1x improvement in user satisfaction during prolonged sessions

Self-Improving Knowledge Graph

Bot trained on every user interaction, tuning answer accuracy over time

Result: 41% reduction in fallback cases over the first 60 days

Our AI/ML Architecture

Core Models

Semantic Retrieval Engine:
- Hybrid BM25 + dense vector search (OpenAI Embeddings + FAISS)
- Surfaces exact-match answers using semantic proximity rather than keywords
Memory-Aware Dialogue Layer:
- GPT-4 with memory state + rule-based topic shift detection
- Handles multi-turn conversations with domain-switching capability
Continuous Learning & Feedback Loop:
- Learns from thumbs up/down, agent escalations, and unresolved queries
- Self-updates response ranking through reinforcement tuning