Selvin PaulRaj KAI Engineer · MCP · Agentic Systems · RAG · LangGraph
3030+Projects1 yr 1 moFull-time
I build AI agents, MCP servers, and LangGraph multi-agent pipelines that run in production. Before agents I shipped full-stack MERN and Next.js apps, and I'm now doing an M.E. in Computer Science (AI).
At Zinnov (Draup)I built DRI-MAS, a LangGraph multi-agent system that turns plain-English requests into validated SQL reports. Analyst turnaround went from hours to under a minute. Outside work I publish MCP servers and CLI agents.
“My agent pipeline took analyst turnaround from hours to under a minute.”
I'm an AI engineer in Chennai, doing a Master's in Computer Science (AI specialisation) at Kings Engineering College after a B.Tech in Information Technology.
I specialize in agentic AI: LangGraph multi‑agent pipelines, MCP servers, RAG systems, and LLM‑powered automation. I test them like production code, with test suites, retrieval metrics, and red‑team cases. The full‑stack side (Next.js, React, Node.js, Python) lets me ship the API and interface around the agent too.
At Zinnov (Draup), I build AI automation for workforce intelligence: a LangGraph multi‑agent system that turns plain‑English requests into validated SQL reports, hybrid FAISS + BM25 search over 2.6M company records, and LLM‑driven role‑mapping pipelines. Outside work, I ship open‑source AI tooling: MCP servers, CLI agents, and full‑stack apps.
Open to AI Engineer roles·Chennai, India · GMT+5:30
SYS · agent.meshLIVEnodes 07 · edges 12v2.4
3030+
Projects Shipped
11yr11mo
Full-time
Intern8 mo8 mo
Certs7171
Now StackLangGraphMCPRAGNext.jsPythonFastAPI
Skills
The stack behind the agents, retrieval systems, and tooling I ship.
01Agentic AI
LangGraph
LangChain
CrewAI
Multi-Agent Systems
Model Context Protocol (MCP)
Agent-to-Agent (A2A)
Tool / Function Calling
Agent Memory
Human-in-the-Loop
Agent Planning & Task Decomposition
02LLMs
Claude API
OpenAI API
Gemini API
Prompt Engineering
Context Engineering
Structured Output (Pydantic)
LLM Evaluation
03RAG & Search
RAG
FAISS
Hybrid Search (BM25 + Vector)
Embeddings
Semantic Search
Re-ranking
Entity Resolution
Agentic RAG
04Evaluation
JEV / Typed Decision Models
Agent Evaluation
LLM-as-a-Judge
RAG Evaluation
Tool-Use Evaluation
Evaluation Harnesses
Ablation Testing
LLM Observability
Tracing & Telemetry
05AI Security
Prompt Injection Defense
AI Guardrails
LLM Red Teaming
PII Detection & Sanitization
06Languages
Python
TypeScript
JavaScript
SQL
07Backend & Data
FastAPI
Node.js
Express.js
PostgreSQL
MongoDB
Trino
Pandas
ETL Pipelines
08Frontend
React
Next.js
Tailwind CSS
Framer Motion
09Cloud & Tooling
AWS (S3, Glue)
Docker
Git & GitHub
pytest
Playwright
Web Scraping
Vercel
Projects
Production agent systems from my work at Draup, plus personal projects, some open source and some kept private.
01Work · Draup
DRI-MAS: Multi-Agent Reporting System
Analysts spent hours turning business questions into SQL and Excel reports. I architected a LangGraph multi-agent system (intent and manager agents, schema router, SQL generator, memory and evaluation layers) that converts plain-English requests into validated Trino SQL and structured reports, with hybrid FAISS + BM25 retrieval over 2.6M company records and human-in-the-loop checks. Turnaround dropped from hours to under a minute.
Single-user personal AI: a rule router and an LLM dispatcher hand each request to one of nine specialist agents with scoped tools (106 in total, each with a server-set risk level), running a six-step function-calling loop with one-level delegation. Agent turns stay on private-safe models (Groq, then local Ollama), every provider sits behind a circuit breaker, consequential actions run only when they match the owner's own words, and sends and deletes wait on an audited confirm card. Also reachable from Telegram and WhatsApp. 1,144 tests.
Job board for Indian freshers: 12 source adapters with robots.txt checks, explainable rule-based classification, a scam shield, and AI resume tools behind a post-generation fact guard that rejects any new number, skill, or name. Every AI feature has a deterministic fallback. About 1,305 tests plus 826 pgTAP assertions across 11 CI workflows.
Weekly business-signal pipeline for 537 enterprise accounts across 8 industry verticals: newsroom discovery and news search, pooled-browser fetching, a confidence-scored date ladder, four-layer deduplication and union-find event clustering, one LLM classification per event validated against the signal taxonomy, analyst review for uncertain events, and deterministic top-2-per-vertical ranking with reconciled CSV/XLSX reports.
Matching company names across datasets broke on acronyms, subsidiaries, and spelling variants. I built an entity-resolution engine over 2.67M company and subsidiary records: FAISS semantic recall, BM25 rescoring on the candidates, acronym and parent-name boosts, and parent-company roll-up, with an LLM decision agent below a 0.85 confidence threshold and a human for genuine ties.
Tags every mapped skill in a job description with one of five proficiency levels, backed by a quote from the JD. Deterministic evidence (years of experience, level phrases and verbs) is used first; remaining skills go to the LLM in batches with a mandatory quote, a substring check, and a second verifier call, and anything unsupported is recorded as not stated instead of guessed.
Auditable accounts-payable workflow that reads text and scanned PDFs (OCR), CSV, or JSON invoices, matches each one against the PO, goods receipt, and vendor master, checks duplicates, payment fraud, and GST errors, and ends in approve, hold with one question for a reviewer, or reject. Rules decide and models only advise, with every check citing its source record. Backed by 79 tests (incl. 20 guardrail tests), 31/31 labelled cases, 32/32 PDF fields, 0 false approvals across 600 adversarial variants, 0 of 489 prompt-injection attacks changing a decision, and an LLM-as-judge validated against a human-labelled gold set, all behind a regression gate in CI.
Single-user AI financial advisor over read-only Kite (Zerodha) data and live web search. A regex safety gate runs before any model, a deterministic router sends questions to LangGraph workflows (with a Google ADK agent only for unmatched questions), Kite MCP tools load via langchain-mcp-adapters behind an allow-list of 16 read-only tools, all numbers are computed in code with provenance tags, and an LLM router falls back from NVIDIA NIM to Gemini Flash with every call logged.
Streamlit lab for comparing RAG techniques side by side: 11 chunking strategies, dense and sparse embeddings, 7 retrieval strategies (keyword, boolean, TF-IDF, BM25, dense, hybrid RRF, MMR), and 6 reranking strategies, scored with Precision@k, Recall@k, MRR, MAP, and NDCG@k plus before/after rerank impact on a built-in labelled set. Each stage is a self-registering plugin registry covered by a 95-test pytest suite.
MCP server that exposes 36 allow-listed Instagram actions (posts, carousels, comments, DMs, insights) to Claude and ChatGPT through Composio, with automatic OAuth, structured JSON logging, health and usage stats, and fail-safe error handling. Deployed on Alpic over streamable HTTP.
Published npm CLI and MCP server for Microsoft Outlook via Microsoft Graph: one-time OAuth, rich terminal output with a --json/--ai mode for agents, and 21 MCP tools covering email, calendar, folders, and rules.
Governance layer for five AI agents sharing one git repository: event-sourced state that changes only when a different party verifies an event, a deterministic reducer checked in CI, tiered permissions with owner-only approvals, and a five-gate security CLI for injection, actions, outbound secrets, commits, and audit. About 441 tests.
Python · Multi-Agent Systems · Event Sourcing
2026
FleetVoice: WhatsApp Agent for Truck Drivers
Hackathon-built WhatsApp agent that turns Tamil, Hindi, and English voice notes into trip updates: a typed Jev classifier first, then Gemini, NVIDIA NIM, and rules, each behind a circuit breaker; an input guard before any model, an output guard that checks numbers and places against the driver's words, no model tools, and a feedback loop that tunes the confidence bar.
Python · FastAPI · LLMs
2026
SPR Studio: AI Video Pipeline
Script-to-video pipeline: a zero-shot clone of my own voice, Whisper word alignment, Rhubarb visemes with lip-sync rules, a seekable HTML comic renderer, and an ffmpeg mix to -14 LUFS in four formats and two styles, with an AI disclosure label in every frame.
Chrome extension that indexes every ChatGPT and Gemini prompt and response in real time: search and jump to any turn, export as Markdown or JSON, or sync conversations to Notion as native blocks.
Sibling of the Instagram server: exposes LinkedIn actions to Claude and ChatGPT with dynamic tool discovery from Composio, automatic OAuth, structured logging, and health reporting.
Company name to official logo: parallel Bing and Serper search, official-URL ranking, then a three-tier extractor (logo.dev API, headless browser for JS-rendered sites, HTML scraping) behind a Streamlit UI.
AI content generator with dynamic templates, searchable filters, adaptive forms, and rich-text editing. Built on Next.js, Gemini, Clerk, Drizzle ORM, and PostgreSQL.
npm CLI that strips a fresh Vite + React project down to a clean starter: removes boilerplate, simplifies the component structure, and prepares Tailwind CSS.
Full-time AI automation work, internships, and study.
Work 06
Associate Data Analyst, AI Automation (current role)
Sep 2025 - Present (1 yr 1 month)
Zinnov (Draup)·Arakkonam, Tamil Nadu, India (On-site)
Architected DRI-MAS, a LangGraph multi-agent system that turns natural-language business requests into validated Trino SQL and structured Excel reports, cutting analyst turnaround from hours to under a minute.
Built its 10 core components (intent and manager agents, normalisation engine, schema router, SQL generator, memory layer, evaluation framework) behind 4 FastAPI endpoints, covered by 30 automated test suites.
Implemented hybrid retrieval (FAISS + BM25) over 2.6M company records with human-in-the-loop validation.
Built a Company Synonym Agent for entity resolution across 2.6M company and subsidiary records: semantic matching, acronym validation, and parent-company resolution.
Automated business-signal monitoring for 537 enterprise accounts across 8 industries through the Deal Origination Framework (DOF) Agent: scraping, deduplication, LLM classification, scoring, and reporting.
Delivered LLM-powered role mapping for enterprise clients, including 746 role–task–skill mappings for taxonomy alignment.
Python
LangGraph
Claude API
FastAPI
Trino
PostgreSQL
FAISS
BM25
AWS S3/Glue
Playwright
Pydantic
pytest
Tech Intern
Jun 2025 - Sep 2025 (3 months)
Zinnov (Draup)·Arakkonam, Tamil Nadu, India (On-site)
Built the RMS Agent, an LLM-powered role-mapping system that standardises job descriptions against Draup's workforce taxonomy using Gemini and FAISS semantic search.
Automated Existing / New / Intern role classification with parallel batch processing.
Built scraper modules (Requests, BeautifulSoup, Playwright, GNews API) with deduplication, the foundation of the later DOF Agent.
Designed the architecture for a 6-stage business-signal intelligence pipeline.
Python
FAISS
Gemini 2.0 Flash
Playwright
BeautifulSoup
Pandas
Python Development Intern
May 2024 - Jul 2024 (2 months)
OCTANET SERVICES PVT LTD·Bhubaneswar, Odisha, India (Remote)
Built FutureFunds, a Flask web app that projects investment outcomes from user inputs.
Implemented the forecasting engine with NumPy, pandas, and TensorFlow, vectorising the hot paths.
Built the responsive frontend with HTML, JavaScript, and Firebase.
Web Development Intern
Mar 2024 - May 2024 (2 months)
VERITECH SOFTWARE IT SERVICES·Chennai, Tamil Nadu, India (Remote)
Built a responsive React portfolio site with an integrated generative-AI chatbot.
Developed a 3D product landing page with Next.js and React-Spline.
Built a MERN expense tracker with category analytics and charts.
Full Stack Intern
Jul 2023 - Aug 2023 (1 month)
SERVIMOS TECHNOLOGIES PVT LTD·Chennai, Tamil Nadu, India (On-site)
Built a MERN-stack Library Management System covering cataloguing, member records, and issue/return tracking.
Restructured React components and state management to remove unnecessary re-renders.
Shipped features through code review with senior developers.
Co-founder (current role)
Mar 2024 - Present (2 yrs 7 months)
GenXRverse·Chennai, Tamil Nadu, India (Hybrid)
Co-founded a student developer collective that delivered 20+ documented, deployed projects for students and clients.
Organised and sponsored a college hackathon and ran hands-on workshops through the Data Science Club.
Built our own MVP and helped teammates land internships through the projects we shipped together.
Education
Aug 2025 - Present (1 yr 2 months)
M.E Computer Science and Engineering
Kings Engineering College
Pursuing a Master's degree with specialization in Artificial Intelligence, Agentic Systems, Machine Learning, and Intelligent Automation. Building advanced research projects involving LLMs and distributed AI agent workflows.
May 2021 - May 2025 (4 yrs)
B.Tech Information Technology
DMI College of Engineering
Graduated with a CGPA of 8.4 while completing multiple AI, web development, and automation projects that strengthened full-stack engineering and problem-solving skills.
Recognition
1 published research paper and 5 awards across hackathons, coding contests and workplace recognition.
Published/Jun 2025/Research Publication
Event Management System for College Symposium and Events with Facial Recognition Image Sharing System
IRJMETSInternational Research Journal of Modernization in Engineering Technology and Science
Proposed and developed a smart Event Management System (EMS) integrated with an AI-powered Facial Recognition-Based Image Sharing System (FRISS). The platform streamlines the complete event lifecycle including registration, QR-based check-ins, scheduling, secure image sharing, and GDPR-compliant consent management.
Abstract
The system leverages Next.js, Convex, Clerk, Razorpay, TensorFlow, and Cloudinary to automate event workflows and deliver personalized image galleries using facial recognition. It demonstrates a scalable, privacy-conscious, and user-centric approach to academic event digitization.
01 · Key Features
✓Online event registration
✓QR code-based check-in
✓Real-time event scheduling
✓Automated facial recognition tagging
✓Personalized image galleries
✓Consent-based privacy workflows
✓Secure image delivery
02 · Stack
Next.jsTypeScriptConvexClerkRazorpayTensorFlowCloudinaryFacial RecognitionQR CodesAI / ML
Certifications
71 certificates from 11 issuers, focused on agentic AI and LLM engineering.
Featured
AnthropicJul 2026
Claude Certified Architect - Foundations
AI System DesignClaude Agent SDKAgentic AI+3
AI EngineeringView ↗
Hugging FaceJul 2026
Hugging Face Agents Course
AI AgentsAgentic AITool Calling+2
AI EngineeringView ↗
AMDJun 2026
LLM Serving Inference with vLLM and MI300X GPUs
vLLMMI300XLLM Inference+1
AI EngineeringView ↗
AMDJun 2026
AI Agents 201: Design to Deployment: A Hands-On Guide to Multi-Agent Systems
Agent2AgentGoogle Agent Development Kit (ADK)Multi-Agent Systems+1
Agentic AIView ↗
AnthropicMay 2026
Claude with the Anthropic API
Anthropic APIClaudeLLM Integration+2
AI EngineeringView ↗
AnthropicSep 2025
Model Context Protocol: Advanced Topics
MCPModel Context ProtocolAI Agents+2
AI EngineeringView ↗
DeepLearning.AIAug 2025
Agentic Knowledge Graph Construction
Knowledge GraphsAI AgentsRAG+1
AI EngineeringView ↗
Jul 2025
Hugging FaceAI Engineering
MCP Course
MCPModel Context ProtocolAI Tools+1
Jul 2025View ↗
Contact
Hiring for AI engineering, or building agents and MCP tooling? Let's talk.
Open to opportunities
Let's build something together
Whether you have a project in mind, want to explore AI engineering collaboration, or just want to connect, I'm always up for a conversation.
