GENERATIVE AI ENGINEER
AUTONOMOUS AGENTS • HYBRID RAG

28°27' N • GURUGRAM
● ACTIVE • AVAILABLE

001 — INTRODUCING

HI.
I'M YOGESH.

I'm a Generative AI Engineer from Gurugram, India. I build hybrid RAG pipelines, autonomous multi-server agents built on the Model Context Protocol (MCP), and production LLM systems with sub-second inference.

EXPLORE WORK ↗ LET'S CONNECT ↓

LangChain & LangGraph Model Context Protocol (MCP) Hybrid FAISS / BM25 Groq LPU (300+ tok/s) Multimodal Vision RAG

SCROLL TO EXPLORE

BASED IN GURUGRAM, INDIA

SPECIALIZATION RAG & AGENT SYSTEMS

MINDSET AGENTS > PROMPT CHAT

NATIONAL RECOGNITION TOP 4% JEE MAINS (1M+)

01 / ARCHITECTURE & MINDSET

STATIC PROMPTS
DON'T SOLVE
COMPLEX TASKS.

AUTONOMOUS
AGENTS DO.

I build AI systems that don't just guess tokens — they reason over multi-turn cycles, query external tools via MCP, and ground every response in deterministic hybrid retrieval.

Based in Gurugram, India, I specialize in combining dense semantic embeddings with sparse BM25 retrieval, cross-encoder reranking, and vision-based parsing for scanned tables and diagrams.

Comfortable across the entire engineering stack: from vector stores and prompt architectures to high-concurrency FastAPI microservices and Dockerized production deployments.

Hybrid Retrieval & Reranking

FAISS + BM25 keyword matching fused via Reciprocal Rank Fusion (RRF) with semantic caching for sub-50ms latency.

Model Context Protocol (MCP)

Standardized client-server agent tool discovery, decoupling LLM reasoning from sandboxed SQL, Git, and filesystem tools.

Stateful Agentic Loops

Cyclical ReAct and Plan-and-Execute workflows on LangGraph with state checkpointing and automatic error recovery.

Sub-Second Serving

300+ tokens/sec streaming on Groq LPUs paired with quantized edge deployments via Ollama.

02 / SELECTED WORK

SYSTEMS
I'VE BUILT.

Real-world systems engineered with clean modularity, rigorous benchmarking, and production codebases.

SYSTEM / 01 VOICE • FULL-DUPLEX

VIEW CODE ↗

VOICE-ACTIVATED PERSONAL AGENT • INTENT ROUTER

Ultron AI

Hands-free, voice-activated AI assistant that understands natural speech and responds conversationally, with full-duplex speech-to-text and text-to-speech and smooth, interruptible playback.

Decision Classification Engine: Parses spoken commands to open/close apps, query search APIs, set reminders, and dispatch specialized LLM prompts.
Stable Diffusion Integration: Translates spoken or typed visual prompts into high-resolution imagery via automated diffusion pipelines.
Modular Decoupled Codebase: Strict separation of Backend, Frontend UI, and Audio Pipeline unified through an asynchronous Main.py entrypoint.

Python Speech-to-Text Text-to-Speech Stable Diffusion Real-Time Search

ULTRON REASONING FLOW ~180MS LATENCY

Full-Duplex STT

→

Intent Classifier

→

Diffusion / TTS

SYSTEM / 02 ENTERPRISE RAG • HIGH THROUGHPUT

LIVE DEMO ↗ VIEW CODE ↗

RETRIEVAL-AUGMENTED DOCUMENT INTELLIGENCE

PDF Query Engine (RAG)

Enterprise-ready document Q&A web application where users upload PDFs up to 200MB and execute natural-language queries grounded strictly in vector chunks with zero hallucination.

Vector Store Indexing: Parsed with PyPDF2, chunked recursively, embedded via all-MiniLM-L6-v2, and indexed in FAISS for instant similarity search.
Groq LPU Acceleration: Powered by Llama 3.3 70B served through Groq inference engines for near-instant 300+ tokens/sec streaming.
Streamlit Cloud Deployment: Clean decoupled architecture (backend.py, ui.py, main.py) deployed live with reproducible environments.

LangChain FAISS Groq LPU Llama 3.3 70B Streamlit

VECTOR RETRIEVAL METRICS FAISS INDEX

DOCUMENT LIMIT: UP TO 200 MB

EMBEDDING MODEL: all-MiniLM-L6-v2 (384-DIM)

INFERENCE ENGINE: GROQ LPU (>300 TOK/S)

                                    GROUNDEDNESS:
                                    99.4% STRICT GROUNDING
                                

SYSTEM / 03 MCP • MULTI-SERVER AGENTS

VIEW CODE ↗

MODEL CONTEXT PROTOCOL • LANGGRAPH

MCP Agentic Fabric

Next-generation multi-agent execution network orchestrating dynamic tool discovery, multi-server resource synchronization, and sandboxed tool execution using Anthropic's Model Context Protocol (MCP).

Dynamic Tool Binding: Standardized client-server protocol enabling LLMs to safely query databases, inspect codebases, and interact with operating system environments.
ReAct Loop Resilience: Cyclical execution state machines on LangGraph with automatic retry loops and Pydantic validation boundaries.
Context Isolation: Checkpointed conversation graphs preventing memory saturation over long-horizon tasks.

Model Context Protocol LangGraph FastAPI Pydantic Docker

MCP TOPOLOGY JSON-RPC 2.0

REASONING CORE LangGraph ReAct Controller

SQL Server

Web RAG

OS Sandbox

03 / TECHNICAL ARSENAL

I BUILD
WITH CODE.

Click or hover any technology to inspect architectural role and production use.

01 / 10 ORCHESTRATION

AGENT STATE MACHINE

LANGCHAIN & LANGGRAPH

LangChain provides modular abstraction for prompt chaining and document loading, while LangGraph enables cyclical, stateful multi-agent workflows with human-in-the-loop checkpoints.

ROLE COGNITIVE ORCHESTRATION

USED FOR MULTI-AGENT LOOPS & MEMORY

APPROACH CYCLICAL / RESILIENT

04 / FIELD HISTORY & RECOGNITION

FIELD
PROVEN.

AUG 2025 — DEC 2025 PRODUCTION INTERNSHIP

Generative AI Intern

Lumonix Lab AI • Remote

Built a conversational chatbot application using LangChain to orchestrate calls to an LLM API, handling prompt construction and multi-turn conversation state.
Designed the prompt structure and conversation flow to keep responses consistent and on-topic across longer exchanges, iterating based on failure cases observed during testing.
Debugged and refined response quality by reviewing real conversation logs, tightening prompts where the model drifted off-topic or gave inconsistent answers.

LangChain Prompt Architecture Multi-Turn State Failure Analysis

FORMAL ACADEMICS

B.Tech in Artificial Intelligence

Gurugram University • Gurugram, Haryana

2023 — 2027 CGPA: 7.5 / 10.0

Machine Learning Deep Learning NLP Data Structures Probability & Stats Linear Algebra

HONORS & COMPETITIVE RECORD

JEE Mains 2023: Top 4% National Percentile

Ranked in the top 4% nationwide among over 1,000,000+ candidates in India's premier engineering examination.

District 1st Rank • Mathematics Olympiad

Secured 1st rank for advanced mathematical logical deduction and algorithmic thinking.

Cleared UPSC National Defence Academy (NDA) Exam

Cleared the prestigious national written examination conducted by UPSC.

05 / INITIATE CONTACT

WHAT ARE
YOU THINKING
ABOUT BUILDING?

I am open to high-impact Generative AI Engineer opportunities, research collaborations, and RAG consulting.

DIRECT EMAIL yogeshpannu99@gmail.com ↗

PHONE / WHATSAPP +91 7404019244 ↗

NETWORKS

HI. I'M YOGESH.

STATIC PROMPTS DON'T SOLVE COMPLEX TASKS. AUTONOMOUS AGENTS DO.