portfolio v2.0

UMAKARTHIKEYA

0%
Available for AI Engineer Roles

UMAKARTHIKEYA

Building Production-Ready |

AI Engineer specializing in Agentic AI, MCP Architectures, RAG Systems, LLM Applications, and Intelligent Automation.

2+
Production AI Systems
4+
Major AI Projects
85%+
ML Model Accuracy
400MB→5MB
Memory Optimized
scroll

// about me

Building AI systems that
solve real problems

I'm an AI Software Engineer with hands-on experience designing and deploying LLM-powered applications and agentic AI systems in production environments.

My work spans the full AI stack — from RAG pipeline design and vector retrieval optimization, to FastAPI backend development and containerized cloud deployments. I've diagnosed production OOM crashes, built multi-system n8n automations, and architected MCP-based tool registries that decouple reasoning from tooling.

Engineering rigor matters as much as AI knowledge.

LangGraphFastMCPLangChainRAGFAISSFastAPIPythonGroqOpenAIDockern8nStreamlitRenderPostgreSQLscikit-learn

Agentic AI Systems

Design multi-node LangGraph pipelines with stateful memory, conditional branching, and production-grade error handling.

MCP Architecture

Implement Model Context Protocol to decouple tool definitions from agent reasoning — enabling dynamic, extensible tool registries.

RAG Pipelines

Architect FAISS-backed retrieval systems with tuned chunk sizes, embedding overlap, and memory-optimized fallback strategies.

Production Deployment

Containerize full-stack AI apps with Docker, deploy to Render, and orchestrate automated workflows via n8n.

// technical skills

Full-stack AI engineering

AI Engineering

Agentic systems, multi-node pipelines, and LLM orchestration.

LangGraphFastMCPMCP ArchitectureRAGAgentic AIFunction CallingVision LLMsLangChain
LLM APIs
Groq APIOpenAI APILlama 3.1/3.3
Backend
PythonFastAPIFlaskDjangoAsync PythonPydanticREST APIs
Vector & Data
FAISSEmbedding RetrievalSemantic SearchPandasNumPyETL
Cloud & DevOps
DockerRenderGitHubLinuxn8nWorkflow Automation
ML & Vision
scikit-learnYOLOv8OpenCVFeature Eng.Hyperparameter Tuning
Databases
PostgreSQLMySQLSQLiteJoinsCTEsWindow Functions

// featured projects

Things I've built & shipped

Each project is a production AI system — not a tutorial, not a demo. Real problems, real architecture decisions, real deployments.

01Agentic AI · Production · RAG

AdOps Intelligence Agent

AI-powered campaign monitoring and optimization platform

Problem

Ad operations teams waste hours manually analyzing campaign KPIs, identifying underperformers, and writing optimization recommendations across dozens of campaigns.

Solution

A fully automated agentic AI pipeline: upload a CSV to Google Drive → n8n triggers analysis → LangGraph orchestrates 3 nodes → Groq LLM reasons over RAG-retrieved strategies → Gmail delivers recommendations in <2 minutes.

<2min
Pipeline Latency
400MB→5MB
Memory Fixed
8+
Workflows
Zero
Manual Intervention
LangGraphFastAPIFAISSGroq Llama 3.1n8nDockerRenderStreamlitPython

// system flow

Google Drive Upload
n8n Webhook Trigger
FastAPI Backend
KPI Analyzer Node
FAISS RAG Retrieval
Groq Llama 3.1 LLM
Gmail Alert System
Architecture
02MCP Architecture · Vision AI · Healthcare

MediBot MCP

Multimodal AI healthcare assistant using Model Context Protocol

Problem

Healthcare information access is fragmented — symptom analysis, medication identification, and multilingual support each require separate tools or expensive proprietary systems.

Solution

A unified multimodal AI agent using FastMCP and Groq Llama 3.3 70B. MCP decouples tool definitions from agent reasoning, enabling dynamic tool selection. Vision LLM identifies medications from photos. gTTS delivers responses in English and Telugu.

70B
Model Size
2 langs
EN + Telugu
Vision
Med Imaging
Dynamic
Tool Selection
FastMCPGroq Llama 3.3 70BVision LLMLangChaingTTSStreamlitRenderPython

// mcp architecture

Streamlit Client
MCP Client Layer
FastMCP Server
Tool Registry (MCP)
Groq Llama 3.3 70B
Vision LLM Module
gTTS Voice (EN/Telugu)
Architecture
03Computer Vision · Accessibility · Offline

Smart Assistive Identifier

Real-time object detection and face recognition — 25+ FPS on CPU

Problem

Assistive technology for visually impaired users is either expensive, requires internet connectivity, or fails to run on modest hardware without GPU acceleration.

Solution

An offline, self-contained executable combining YOLOv8 object detection and face_recognition library at 25+ FPS on CPU only. Confidence thresholds and frame sampling rates were tuned for throughput. Multilingual voice alerts in Telugu and English via pyttsx3. Dark-mode Tkinter GUI.

25+ FPS
Real-Time
CPU only
No GPU
Offline
No Internet
2 langs
Voice Output
YOLOv8OpenCVface_recognitionpyttsx3TkinterPython

// pipeline

Camera Input (OpenCV)
YOLOv8 Detection
face_recognition Library
CPU Optimized (25+ FPS)
pyttsx3 Voice Engine
Tkinter Dark-Mode GUI
Architecture
04ML · Full-Stack · Clinical

Heart Disease Prediction

Django ML web app with 88%+ cross-validated accuracy

Problem

Clinical decision support tools are either too complex for small clinics or lack the transparency needed for medical use — black-box models with no explainability.

Solution

A Django-backed ML web application with full EDA → feature selection → ensemble modeling (Logistic Regression, KNN, Random Forest) → REST API pipeline. 88%+ cross-validated accuracy on clinical datasets.

88%+
CV Accuracy
3 Models
Ensemble
REST
API Layer
Full
ML Pipeline
scikit-learnLogistic RegressionRandom ForestKNNDjangoPandasPython

// pipeline

Clinical Dataset Input
EDA + Feature Selection
Logistic Regression
KNN + Random Forest
88%+ CV Accuracy
Django REST Endpoint
Architecture

// work experience

Where I've worked

AI & ML Engineer Trainee

Innomatics Research Labs

Hyderabad, India

May 2025 – May 2026

Designed and deployed structured LLM workflows for document summarization, entity extraction, and automated reporting — implementing Pydantic output validation schemas to enforce response reliability across production AI pipelines.

LLMsPrompt EngineeringPydanticOutput Validation

Built and evaluated RAG (Retrieval-Augmented Generation) pipelines for domain-specific Q&A over structured datasets — combining embedding-based retrieval with LLM reasoning to reduce hallucinations and improve answer accuracy.

RAGEmbeddingsVector SearchLLM Reasoning

Developed agentic AI task automation scripts using LLM function-calling and tool-use patterns — enabling autonomous multi-step execution across data ingestion, transformation, and reporting workflows with zero manual intervention.

Agentic AIFunction CallingTool UseAutomation

Operationalized LLM and ML model outputs as FastAPI REST endpoints with async request handling, Pydantic schema validation, and structured error responses integrating AI inference into downstream business pipelines.

FastAPIAsync PythonREST APIsML Serving

Trained, fine-tuned, and cross-validated ML and DL models (classification, regression, clustering, neural networks) through systematic hyperparameter tuning and feature engineering achieving 85%+ accuracy on held-out validation sets.

MLDeep Learningscikit-learnHyperparameter Tuning

Designed ETL pipelines and feature engineering workflows (Pandas, NumPy) to prepare high-quality training data for LLM fine-tuning and ML model development reducing manual data processing effort by 50%.

ETLFeature EngineeringData PipelinesLLM Data Prep

// education

B.Tech — Computer Science & Engineering

AI & ML Specialization

SVIT (JNTUH), Hyderabad · 2021–2025

ML · DSA · OS · DBMS · Software Engineering · Computer Networks

// certifications

Data Science with AI/ML, GenAI & Data Analytics

Innomatics Research Labs · 120+ hrs · 2025

Machine Learning Internship with Python

Verzeo Edutech

MERN Full Stack Web Development

NxtWave · In Progress

// let's connect

Let's Build the
Future of AI
Together.

AI EngineerApplied AI EngineerAgentic AI DeveloperGenAI EngineerML Engineer

Actively seeking roles where engineering rigor matters as much as AI knowledge. Open to remote and on-site opportunities.