import tensorflow as tf
model = tf.keras.Sequential([tf.keras.layers.Dense(128, activation='relu')])
def train_neural_network(X_train, y_train):
optimizer = tf.keras.optimizers.Adam(learning_rate=0.001)
model.compile(optimizer=optimizer, loss='categorical_crossentropy')
from transformers import AutoModel, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('bert-base-uncased')
import torch
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
class NeuralNetwork(nn.Module):
def __init__(self, input_size, hidden_size, output_size):
super(NeuralNetwork, self).__init__()
self.fc1 = nn.Linear(input_size, hidden_size)
self.relu = nn.ReLU()
self.fc2 = nn.Linear(hidden_size, output_size)
import pandas as pd
df = pd.read_csv('dataset.csv')
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)
model.fit(X_train, y_train, epochs=50, batch_size=32)
accuracy = model.evaluate(X_test, y_test)
Hello, I'm Karthik
Work Experience
AI Engineer Intern
Pronic Inc
Remote
Oct 2025 - Present
Key Achievements & Responsibilities
- Contributed as part of the CBE Companies chatbot development team, designing and implementing HR assistant use cases on the Kore.ai Agentic Platform, including Profile Management and Absence Management, with seamless integration to Workday APIs for real-time employee data retrieval.
- Collaborated on end-to-end dialog design, intent logic, and workflow automation to support internal HR self-service operations, ensuring functional accuracy and smooth enterprise integration.
- Validated conversation accuracy, response precision, and flow reliability on the JBL project through executing structured test scenarios as part of the QA/testing team on the Kore XO Platform, ensuring high-quality interactions.
- Improved bot intelligence by using RAG-based Search AI, extractive NLP, and Generative AI (OpenAI/Azure), resulting in enhanced retrieval relevance and better conversation quality through iterative analytics-driven tuning.
RAG
NLP
OpenAI
Azure
Workday API
Generative AI
Key Achievements & Responsibilities
- Improved document processing efficiency by developing Python-based workflows to parse and chunk mortgage files, enabling structured search and classification.
- Increased data extraction accuracy by evaluating and integrating OCR engines like Tesseract and EasyOCR, enhancing field-level capture in scanned documents.
- Enhanced AI performance by implementing a RAG pipeline with LlamaIndex, FAISS vector search, and comparing LLMs such as Gemini and Mistral for document Q&A and classification.
- Currently developing a Gradio web interface to provide an intuitive user experience for the AI-powered document processing application.
- Honored as a Top Performer for delivering high-impact AI automation solutions and building production-ready workflow systems.
Python
OCR
RAG
LlamaIndex
FAISS
Gradio
LLMs
Volunteering
ISS
Volunteer
International Student Services Volunteer
University at Buffalo
Buffalo, NY • Aug 2024
- ▸ Welcomed new international graduate students during Fall 2024 orientation at the University at Buffalo
- ▸ Assisted with orientation activities and provided guidance on campus resources
- ▸ Helped create a supportive environment to ease their transition into the UB community
Student Support
Event Coordination
Community Building
HF
Coordinator
HackFiesta - Official Hackathon Coordinator
Mahaveer Institute of Science and Technology, Hyderabad
Hyderabad, India • TechnoMist 2K23
- ▸ Coordinated event planning and logistics for a large-scale hackathon with multiple teams and mentors
- ▸ Managed team coordination and ensured seamless communication between participants and organizers
- ▸ Successfully executed the event with high participation rates and positive feedback from attendees
Event Management
Team Leadership
Project Coordination
Education
University at Buffalo
Courses Taken:
- Machine Learning
- Deep Learning
- Data Visualization Using Tableau
- Computer Vision and Image Processing
Jan 2024 - June 2025
Master's in Computer Science - AI/ML (GPA: 3.7/4.0)
Focusing on advanced machine learning techniques, deep learning, and artificial intelligence applications. Working on real-time AI/ML projects with practical implementations.
Jawaharlal Technological University, Hyderabad
Courses Taken:
- C Programming
- Data Structures
- Database Management System
- Algorithms and Analysis
Aug 2019 - July 2023
Bachelor's in Information Technology
Gained strong foundations in computer science, programming, data structures, and algorithms. Developed initial interest in machine learning and artificial intelligence.
Projects
DocuQuery AI Assistant
A production-ready RAG assistant that lets users query complex documents and get fast, low-latency, context-aware answers.
Stack: LangGraph · Gemini · FAISS/Pinecone · RAG · Hybrid search · Hugging Face Spaces
Key Features:
- Built and deployed a production RAG system (LangGraph + Gemini + FAISS/Pinecone) that reduced document analysis time by ~75% while maintaining low-latency inference (1.5–4.0s).
- Improved retrieval precision by ~30% by architecting a hybrid search pipeline (semantic + BM25 + TF-IDF + cross-encoder reranking), decreasing irrelevant context and increasing answer relevance consistency.
- Delivered end-to-end user workflows with document preview (full-text and chunk-level), cross-document comparison, and one-click export (TXT/PDF/DOCX), increasing task completion efficiency by ~40%.
- Stabilized and productionized deployment on Hugging Face Spaces by implementing robust error handling and fallback logic, reducing startup crashes and runtime failures by ~90%.
Amazon Stock Forecasting Dashboard
An interactive dashboard using LSTM models and technical indicators to forecast Amazon stock trends.
Stack: Streamlit · Python · LSTM/Seq2Seq · RSI · MACD · Bollinger Bands
Key Features:
- Developed a real-time stock forecasting dashboard using Linear Regression, LSTM, and Seq2Seq LSTM models to assist traders in analyzing Amazon (AMZN) stock trends.
- Integrated technical indicators (RSI, MACD, Bollinger Bands) across 6,300+ daily records (2000–2025), enhancing visibility into trading signals.
- Reduced forecasting error by 49% over baseline (MAPE improved from 17.45% to 8.86%) using Seq2Seq LSTM.
- Incorporated sentiment analysis from Yahoo Finance and implemented backtesting to evaluate model effectiveness in real-time environments.
Human Action Recognition Using CNN Models
A deep learning system to classify human actions in images.
Stack: TensorFlow/Keras · VGG/ResNet · GoogLeNet/DenseNet (TL) · Streamlit
Key Features:
- Designed a human action classification system using CNN models — implemented VGG and ResNet architectures from scratch, and applied transfer learning with pre-trained GoogLeNet and DenseNet models for performance optimization.
- Collaborated with a teammate to train and evaluate models on the Stanford 40 dataset (9,500+ labeled images across 40 categories); ResNet50 achieved the highest accuracy.
- Achieved 80% real-time classification accuracy, demonstrating potential for use in surveillance, fitness tracking, or gesture-based interfaces.
- Deployed results via Streamlit to enhance model transparency and aid stakeholder interpretation.
Text Summarization using BART Transformer
An NLP pipeline utilizing the BART model to summarize long documents into concise, coherent summaries.
Stack: PyTorch · Hugging Face · BART · Streamlit
Key Features:
- Fine-tuned the BART model on the CNN/DailyMail dataset for abstractive text summarization.
- Achieved a ROUGE-L score of 0.41, indicating high-quality summary generation.
- Deployed the model using Streamlit for real-time summarization capabilities.
- Enabled users to input lengthy articles and receive concise, coherent summaries instantly.
Predicting Hospital Inpatient Costs Using Machine Learning
A machine learning system to predict hospital inpatient costs.
Stack: Python · scikit-learn · Random Forest/GBM/Linear · Pandas · Feature engineering
Key Features:
- Developed a regression pipeline to estimate inpatient costs from a 50,000-record anonymized healthcare dataset.
- Engineered features through imputation, outlier removal, and encoding, reducing nulls by 100% and improving data quality.
- Benchmarked Linear Regression, Gradient Boosting, and Random Forest, achieving R² = 0.82 with the best-performing model.
- Visualized key cost drivers to improve insight accessibility for healthcare analysts and decision-makers.
Sentiment Analysis Using LSTM Architectures
A deep learning pipeline to classify sentiments in text data.
Stack: TensorFlow/Keras · LSTM/BiLSTM · NLP preprocessing · Regularization/tuning
Key Features:
- Constructed an NLP pipeline for sentiment analysis on an anonymized news headlines dataset, applying tokenization, lemmatization, and a 64/16/20 train-validation-test split for robust evaluation.
- Trained multiple LSTM architectures, including a 3-layer LSTM with 86% test accuracy and a Bidirectional LSTM achieving over 92% accuracy.
- Enhanced model performance through hyperparameter tuning and dropout regularization, resulting in a 12% improvement in F1-score and preventing overfitting.
- Streamlined model experimentation and validation workflows, reducing update time by 27% and enabling timely deployment of the best-performing model.
Skills
Currently Learning
Continuously expanding my knowledge and skills
Terraform
Infrastructure as Code
Databricks
Data Analytics Platform
BigQuery
Data Warehouse
Certifications
Get In Touch
Let’s build something meaningful with AI and data.