Back to portfolio

NikhilGPT — Personal AI Knowledge Agent

A RAG-powered professional AI assistant that converts my professional experience, project portfolio, achievements, and AI initiatives into an intelligent, searchable assistant.

RAGGPT-4o-miniOpenAI EmbeddingsPineconen8nGoogle DriveRedisGoogle SheetsLovable

NikhilGPT

Online · Grounded in source docs

Tell me about Nikhil's banking transformation work.
Nikhil led 50+ enterprise workflow transformations for Tier-1 banks — cutting loan TAT from 5 days to ~1 day and freeing 9 FTEs through compliance automation.
Ask about experience, projects, AI work…
Project Overview

NikhilGPT is a personal AI knowledge agent designed to convert Nikhil's professional experience, project portfolio, achievements, BFSI domain expertise, and AI initiatives into an intelligent searchable assistant.

The objective was to build a practical AI system that can answer questions about Nikhil's professional background in a structured and scalable way while staying grounded in curated source documents.

Business Objectives

Why NikhilGPT exists

Personal Knowledge Retrieval

Answer questions about Nikhil's background, projects, achievements, banking transformation work, technical expertise, and AI initiatives.

Interactive Professional Showcase

Move beyond a static resume by providing an AI-powered professional interface.

Applied AI Implementation

Demonstrate practical implementation of RAG, embeddings, vector search, prompt engineering, session memory, fallback handling, and evaluation automation.

Key Use Cases

What it can do

Answer questions about professional profile

Retrieve project achievements and case studies

Summarize banking transformation experience

Present technical and leadership capabilities

Explain AI initiatives

Handle greetings and assistant capability questions

Escalate unanswered questions using fallback email capture

High-Level Architecture

End-to-end request flow

From user query to grounded response — with vector search, session memory, and fallback escalation.

User01
Lovable Web Interface02
n8n Webhook03
Query Optimization04
Pinecone Vector Search05
Retrieved Context06
Redis Session Memory07
GPT-4o-mini08
Grounded Response09
Fallback Email Flow (if needed)10
Architecture Layers

Six engineered layers

01 Knowledge Ingestion Layer

  • Google Drive stores source documents
  • n8n detects new or updated files
  • Text is extracted and prepared for retrieval
  • text-embedding-3-small generates embeddings
  • Pinecone stores vectorized chunks

02 Query and Retrieval Layer

  • User submits question through web interface
  • n8n receives query through webhook
  • Retrieval-optimized query is generated
  • Pinecone retrieves relevant chunks

03 Response Generation Layer

  • GPT-4o-mini receives user query, memory, and retrieved context
  • Response is generated as Nikhil's AI Assistant
  • Answers remain grounded in curated documents

04 Session Memory Layer

  • Redis stores recent conversation turns per active session
  • Enables follow-up questions like “tell me more” and “what was the impact?”

05 Fallback Handling Layer

  • If relevant information is not available, the assistant asks for email
  • Question is forwarded to Nikhil
  • User receives confirmation

06 Evaluation Layer

  • Google Sheets stores structured test questions
  • n8n runs automated evaluation
  • Outputs are captured for quality review
AI Concepts Applied

The applied AI stack

Retrieval-Augmented Generation

Embeddings

Vector Search

Prompt Engineering

Session-Based Conversation Memory

Fallback-to-Human Workflow

LLM Evaluation

MVP Results

Measured outcomes

100

Test questions

80

Answered normally

20

Fallback responses

80%

Overall answer rate

100%

Boundary compliance

88.9%

Non-boundary answer coverage

Challenges Solved

Problems tackled along the way

Cluttered RAG source data

Restructured source documents and reduced duplicated information.

Binary response behavior

Refined prompts to improve intent handling and reduce unnecessary fallbacks.

Lack of conversational continuity

Added Redis session memory.

Weak visibility into performance

Created automated evaluation workflow with Google Sheets.

Project Impact

NikhilGPT demonstrates how RAG, vector search, workflow orchestration, session memory, fallback escalation, and evaluation can be combined into a working professional AI assistant.

Future Enhancements

What's next

Improve recruiter-style response synthesis

Reduce false fallback on valid questions

Improve metadata and chunking strategy

Add better conversation state handling

Enhance evaluation scoring

Refine UI/UX

Ask NikhilGPT about my experience, projects, and AI work