Problem
Support teams were overloaded with repetitive L1 questions, while scripted chatbots gave poor answers and earned little customer trust.
Solution
I built SmartChat: a multi-tenant SaaS platform where businesses upload docs, generate grounded knowledge indexes, and deploy branded AI chat assistants across channels.
My Role
Lead Full-Stack AI Engineer, owning architecture through delivery.
- Architected the end-to-end RAG pipeline and tenant isolation model
- Implemented ingestion, chunking, embeddings, retrieval, and response orchestration
- Integrated AWS document intelligence for OCR-heavy files
- Built analytics surfaces for conversation quality and deflection
Technical Decisions
- Hybrid LLM strategy (OpenAI + open-source) for per-tenant quality and cost control
- Tenant-scoped vector indexes for data isolation and query precision
- Retrieval-first response policy to minimize hallucinations
- Async ingestion workers for reliable large-document processing
Impact
- Automated a large share of repetitive support intents in pilot accounts
- Reduced median first-response time from minutes to seconds
- Improved agent productivity by deflecting L1 tickets and surfacing unresolved intents
- Enabled faster onboarding via self-serve embed, public link, and API deployment
Key Features
- Hybrid RAG pipeline — Uses LangChain to orchestrate OpenAI and open-source LLMs for context-aware, brand-aligned responses.
- Secure document intelligence — AWS Textract and Rekognition for OCR and image analysis on complex PDFs and diagrams.
- Real-time analytics — Tracks user intent, chat volume, and resolution rates for actionable support insights.
- Omnichannel deployment — One-click web embeds, standalone public links, and REST APIs.
- Enterprise security — Secure file handling with AWS S3 and encrypted metadata storage.
Tech Stack
- Frontend: Next.js, Tailwind CSS, TypeScript
- Backend: Node.js, LangChain
- AI & Data: OpenAI API, Pinecone (vector DB), open-source LLMs (Llama/Mistral)
- Cloud Infrastructure: AWS (S3, Textract, Rekognition, Lambda)