Senior Executive at Nvest Solution | AI/ML & Generative AI

Aditya Naranje

AI/ML engineer buildingRAG systems, LLM agents, MCP tool servers, production AI on AWS

Two years designing, deploying and evaluating RAG systems, LLM agents and multi-agent workflows with LangGraph, MCP and AWS, running in production with measured results.

See the case studies Resume

AI/ML projects shipped: 5+
Articles published on Medium: 8+
RAG & GenAI workflows built: 10+
AI/ML technologies used: 15+

I translate business problems into AI systems with measurable gains in automation, reliability and efficiency.

Aditya Naranje

Thane, Maharashtra, India

Open to AI/ML roles

B.E. in Computer Science with a 9.14 CGPA
Six data and ML internships before graduating
Writes about RAG, agents and ML on Medium
Docker One-Shot repo: 16 stars on GitHub

I'm Aditya Naranje, an AI/ML engineer with two years of experience designing, deploying and evaluating production AI. I build LLM agents, multi-agent RAG services and MCP tool integrations for AI workflows, with multi-tenant data scoping, human-in-the-loop review and rigorous model evaluation.

My toolkit spans LangGraph, LangChain, MCP and vector search (pgvector, Pinecone, FAISS), Python and AWS (Bedrock, Lambda, EC2, SageMaker), plus the ML layer underneath: CatBoost and scikit-learn, Optuna tuning and champion/challenger promotion.

Production RAG

Multi-tenant retrieval with hybrid BM25 + vector search, embeddings and evaluation.

LLM and multi-agent systems

LangGraph orchestration, tool calling, MCP servers and human-in-the-loop review.

ML evaluation and monitoring

Champion/challenger promotion gated on AUC and F1 for churn, propensity and lead scoring.

Cloud and CI/CD

AWS deployment, Docker, GitHub Actions, FastAPI/Flask and serverless automations.

How I work

Measure it
Every system ships with a before-and-after number: review effort, setup time, relevance, hallucination rate.
Keep humans on the writes
Agents get read-only access by default. Anything that changes production waits for a person.
Evaluate every change
Retrieval is scored, models are gated on AUC and F1, and a challenger only wins if it's better.

Building production AI: 2+ yrs
Less manual PR review: 85%
RAG bots in production: 10+
B.E. CGPA: 9.14/10

The path a query takes through the RAG and agent systems below. Scroll to follow it.

01User queryChat UI, API
A question arrives
A user asks in plain language, inside a bot that belongs to one client and one knowledge base.
02AI router
03Retrieval
04RAG context
05LLM
06Tools & agents
07Response

The path a query takes through the RAG and agent systems below.

01
User query
A user asks in plain language, inside a bot that belongs to one client and one knowledge base.
Chat UIAPI
02
AI router
A LangGraph router decides whether the question needs documents, a tool, or both, and which bot's context it may touch.
LangGraph
03
Retrieval
BM25 and vector search run together over pgvector or an isolated Pinecone namespace, so a bot only ever sees its own data.
pgvectorPineconeFAISSBM25
04
RAG context
Documents parsed with Docling and embedded with Voyage AI are assembled into a prompt the model must answer from.
DoclingVoyage AIEmbeddings
05
LLM
An LLM on AWS Bedrock or Groq generates the answer, with structured output and guardrails.
AWS BedrockClaudeLlama 3.3
06
Tools & agents
When data lives in a database, MCP tools run read-only queries. Anything that writes waits for human review.
MCPTool callingHuman-in-the-loop
07
Response
An evaluation loop scores relevance and retries up to three times. In production this cut hallucinations by about 30%.
LangSmithRelevance scoring

Four production and open-source systems, each with the problem, the architecture and the number it moved.

Case study 01Generative AIProduction platform

Isolated multi-bot RAG system

Problem: Ten-plus chatbots share one platform. None may leak another's knowledge, and every answer has to stay grounded.
Solution: Each bot gets its own Pinecone namespace. A LangGraph router sends each query down the right path, and an LLM evaluation loop checks relevance before anything is returned.

bots in production: 10+
better response relevance: ~35%
fewer hallucinations: ~30%

LangGraphPineconeOpenAIDockerAWS EC2Nginx

View the code

github.com/adityanaranje/ISOLATED-RAG-CHATBOT

Architecture

01User queryper bot
02LangGraph routerpicks the path
03Isolated namespacePinecone
04Retrieval + eval3-retry scoring
05Answergrounded

Engineering details

Bot-level context separation with isolated Pinecone namespaces.
Relevance is scored by an LLM validation loop that retries up to three times.
Runs on AWS EC2 with Docker and Nginx; Watchtower rolls out updates automatically.

Case study 02Agentic AINvest Solution

CodeGuard, an automated PR review agent

Problem: Every pull request waited on a full manual review, which slowed down every merge.
Solution: An LLM agent built with LangGraph and Flask reads each pull request, scores issues by severity and posts inline fix suggestions, so reviewers start from a ranked list.

less manual review effort: ~85%
faster PR turnaround: ~70%

LangGraphFlaskLLMsGitHub Actions

View the code

github.com/adityanaranje/CodeGuard

CodeGuard, an automated PR review agent interface

Architecture

01Pull requestGitHub
02LangGraph agentreads the diff
03Severity scoringranks issues
04Inline fixeson the lines
05Merge gateGitHub Actions

Engineering details

Severity scores put critical issues at the top of every review.
Fix suggestions are posted inline, on the exact lines they refer to.
Runs as part of the merge flow through GitHub Actions.

Case study 03Agentic AIPersonal project

AI/ML Job Alerts Pipeline

Problem: Finding relevant AI/ML jobs and manually checking each posting against a resume was time-consuming and made it easy to miss suitable opportunities.
Solution: An automated pipeline fetches AI/ML job postings, uses an LLM chain to evaluate resume–job compatibility, stores structured results in Google Sheets, and sends relevant unsent jobs through Telegram.

automated job monitoring: Daily
resume–job matching: LLM

PythonLangChainLLMSerpAPIGoogle Sheets APITelegram Bot APIGitHub ActionsSQL

View the code

github.com/adityanaranje/Job-Alerts-Telegram-Bot-Airflow

Architecture

01Job Search APISerpAPI
02Job Pipelinefetch & clean
03LLM Matchingresume analysis
04Google Sheetsstore & track
05Telegram Botjob alerts

Engineering details

Built an automated pipeline for AI/ML Engineer, Data Scientist and Generative AI job discovery.
Implemented LLM-based resume–job matching to extract structured details and generate match scores.
Added job deduplication using unique job_id and alert tracking to prevent repeated notifications.
Implemented token optimization by cleaning and shortening job descriptions before LLM processing.
Automated daily execution using GitHub Actions with an Airflow-style DAG pipeline architecture.

Case study 04Agentic AINvest Solution, internal

Product onboarding MCP server

Problem: Setting up a new product took 3–4 days of manual discovery, validation, diagnostics and SQL writing.
Solution: A local MCP server gives AI agents (Claude, Codex, Antigravity) a tool for each onboarding step. Database access is read-only and a person reviews every change.

less setup time: ~95%
instead of 3–4 days: 2–3 h

MCPPythonTool callingSQLHuman-in-the-loop

Internal system, code not public

System design

Screenshots aren't public for this internal system, so here is how it fits together.

Architecture

01Product specinput
02AI agentClaude / Codex
03MCP toolsdiscover, validate
04Read-only DBdiagnostics, SQL
05Human reviewapproves writes

Engineering details

Tools cover discovery, validation, diagnostics and SQL generation.
Read-only database access; generated SQL is reviewed before it runs.
Works with several agent clients: Claude, Codex and Antigravity.

Case study 05Generative AIOpen source

RAG chatbot on Amazon Bedrock AgentCore

Problem: A knowledge chatbot needed accurate retrieval, memory across a conversation, and managed cloud deployment.
Solution: FAISS retrieval over 500+ knowledge chunks, Groq Llama-3.3-70B for answers and thread-based memory, containerised and shipped with the AgentCore CLI to Amazon ECR and the Bedrock runtime.

relevance gain: ~40%
knowledge chunks: 500+

AWS BedrockAgentCoreFAISSGroq Llama-3.3LangChainAmazon ECR

View the code

github.com/adityanaranje/RAG-Chatbot-AWS-Bedrock-Agentcore

RAG chatbot on Amazon Bedrock AgentCore interface

Architecture

01Questionthread
02FAISS500+ chunks
03Bedrock / GroqLlama-3.3-70B
04Memoryper thread
05AnswerAgentCore

Engineering details

FAISS index over 500+ knowledge chunks.
Persistent, thread-based conversation memory.
Containerised and deployed via the AgentCore CLI to Amazon ECR and the Bedrock runtime.

More projects

Agents and generative AI first, then machine learning and data analysis.

Agentic AI

AI email job agent for Telegram

A serverless agent that watches Gmail over OAuth2, classifies emails, extracts job details with LangChain and OpenAI, and pushes alerts to Telegram every 15 minutes from GitHub Actions.

LangChainOpenAIGmail APIGitHub Actions

Aditya Naranje

Production RAG

LLM and multi-agent systems

ML evaluation and monitoring

Cloud and CI/CD

How I work

How my systems answer a question

How my systems answer a question

Isolated multi-bot RAG system

CodeGuard, an automated PR review agent

AI/ML Job Alerts Pipeline

Product onboarding MCP server

RAG chatbot on Amazon Bedrock AgentCore

More projects

AI email job agent for Telegram

LangGraph customer support multi-agent

Agent Builder research agent

GitHub events notifier

WhatsApp shop agent

RAGXRay

NanoChat, a small language model from scratch

LLM fine-tuning with Unsloth

CODEVO repository Q&A

Fitness chatbot

SparkLine

Story generator

Movie recommendation system

Book recommender

IPL win probability predictor

T20 score prediction

Disease prediction web app

Store sales prediction

Flight fare prediction

Face and object detection

British Airways review analysis

T20 cricket data analysis

Onyx Data DNA challenges

Indian crop analysis with SQL

Google App Store analysis

Senior Executive

ServiceNow Internship Trainee

Self-directed study, AI/ML and generative AI

RAG

B.E. Computer Science & Engineering

Higher Secondary (XII)

Secondary School (X)

Key credentials

Designing production-grade agentic AI systems, explained through a restaurant

Let's build something intelligent.