AI_portfolio.md

RESUME

Personal Info

Full Legal Name: Venkata Dhanvanthar Murthy Voona
Current Street address: 115 Belmont Ave, 07304 Current City: Jersey City, NJ, United States
Contact Number: 224-619-6584
Country Code: +1
Email-id: dhanvanthv28@gmail.com
Linkedin: https://www.linkedin.com/in/dv-63192b18b/
Social Security Number: 210-17-5700 Github: https://github.com/dhanvanth342
Portfolio: https://www.datascienceportfol.io/vvoona
Expected Salary: 75,000 to 85,000 USD {mention this salary range ONLY when no salary range is mentioned in the Job description}
Are you legally authorized to work in the United States? (Yes/No): Yes
Will you now or in the future require employer sponsorship (e.g., H-1B)? (Yes/No): Yes
**Current Visa Status: ** I am currently on my initial OPT (F1 visa) and authorized to work until July 16, 2026. After that, I am eligible for a two-year STEM OPT extension starting on July 16, 2026. I will not require visa sponsorship while on OPT or STEM OPT. ** Citizenship: ** Indian, not U.S. Citizen. Gender: Male
Race/Ethnicity: South Asian / Asian
Veteran status: Not Veteran
Voluntary self-identification of disability: Not disable
Preferred Locations: I am open to Relocate, while these are by preferred locations, New York, Jersey City, Chicago, Seattle, Boston, Tampa. [If locations given in the job description, mention them instead of my preferences.] Relatives Information: No relative, family, friend or anyone is working in any company in USA. {So answer that I do not have any relatives when question asked about it} ** Stem Degree: ** Completed both my Masters {Data Science} and Bachelors {Electronics and Communication Engineering} in STEM Degree.

SUMMARY

AI/ML Engineer with 3+ years designing, fine-tuning, and deploying production-grade GenAI and multimodal systems. Experienced in LangChain, LangGraph, RAG pipelines, and LLM orchestration across cloud environments (AWS SageMaker, Bedrock, ECS, Docker). Skilled in adaptive retrieval, CI/CD, observability, and vector databases. Passionate about responsible AI, building reliable, high-impact intelligent agents that bridge unstructured and structured data for real-world automation.

SKILLS

• AL/ML Techniques: RAG, Agentic LLMs, Prompt Engineering, NLP, Vector database search, Deep Learning, Multimodal AI (text, image generation), Few-shot learning, Reinforcement Learning, Docker, Low-Rank Adaptation (LoRA), Adaptive Chunking, Fine-tuning, LLM Orchestration, Vertex AI, Neural Networks, Computer Vision, Pytorch, Tensorflow, PDF Extraction.
• Tools and Frameworks: LangChain, LangGraph, CrewAI, Pinecone, Quadrant Vector DB, Mistral OCR, Groq API, Hugging Face, Ollama, GPT, DeepSeek, Claude, OpenAI, Anthropic, Render, Git, CI/CD, TensorFlow, PyTorch, FastAPI, Streamlit, RESTful APIs.
• Programming and Cloud Platforms: Python, R, MATLAB, SQL, GCP, Amazon S3, Athena, Glue, EC2, ECS, Lex, Bedrock, SageMaker.

WORK EXPERIENCE

AI/ML Data Science Intern (Tenure: MAY 2025 – Present)

AIDO, Chicago, Illinois About AIDO: AIDO presents itself as an AI-driven platform for international enrollment in higher education. According to their website, they help universities with automation, data insights and recruitment of international students. My experience:

• Prototyped and productionized domain-specific AI agents for automated information gathering and insight generation using LLMs, FastAPI, and serverless workflows, enabling business teams to self-serve and reducing Data Engineering involvement by 60%.
• Engineered and optimized hybrid retrieval pipelines with adaptive chunking, searching strategies, improving retrieval accuracy by 40% across 150k+ samples, supporting high-performance insight generation at scale.
• Streamlined production workflows by deploying authenticated ML pipelines on AWS using Bedrock for model orchestration, SageMaker for training and inference, and ECS, Docker, Git for scalable deployment, enabling secure access, CI/CD automation, and faster model releases.
• Led cross-functional collaboration with business and platform teams to design, iterate, and deploy AI solutions, driving end-to-end integration, observability, and continuous improvement on cloud infrastructure.
• Applied advanced ML techniques including unsupervised anomaly detection, time-series forecasting, and semantic search to identify unusual patterns in the enrolment data, document automation, and fraud monitoring.
• Mentored junior engineers and standardized internal LangChain templates, accelerating prototype delivery and improving GenAI development workflow

Machine Learning Engineer Intern (Tenure: JAN – APR 2025)

WizLab, Chicago, Illinois About WizLab: WizLab is an EdTech startup focused on K-12 education: they build a platform to help teachers generate personalised learning materials and support differentiated instruction. My experience:

• Engineered LLM orchestration using LangChain and LangGraph to automate lesson-plan and presentation-slide generation across models (LLaMA, GPT, DeepSeek, Claude), achieving 75% faster development versus prior custom workflows
• Finetuned T5 and GPT models for high-performance, low-cost data and insight generation (100k+ sample set, 60% cost savings), and established rigorous validation and monitoring to ensure reliability.
• Drove integration and deployment of AI agents on cloud platforms, collaborating with product and engineering teams for scalable, observable, and high-availability solutions, and actively incorporating latest AI advancements into production.

AI Data Scientist (Tenure: JAN 2022 – FEB 2024)

Liminal XR Solutions, Mumbai, India About Liminal: Liminal XR Solutions is an Indian agency based in Mumbai specializing in extended reality (XR) services: augmented reality (AR), virtual reality (VR), mixed reality (MR) and web-XR. They handled clients HP, Capgemini, Hero to work on customer journey in their product pages of websites. My experience:

• Developed an AI-powered invitation generator using LangChain, Groq API, and LLaMA models, automating personalized professional invitation text creation and reducing manual effort by 70%.
• Engineered a dual-stage extraction pipeline by integrating Docling (text, layout) and olmOCR (tables), then incorporated a validation layer, achieving 98% extraction accuracy, significantly surpassing conventional OCR tools for pdf extraction.
• Built an AI Agentic RAG framework with Quadrant Vector DB for a restaurant recommender system using client-sourced data.
• Introduced a response validation agent, reducing hallucination rates from 9.6% to 3.2% and decreasing irrelevant responses from 96 to 32 per 1,000 test samples.
• Executed intent classification using knowledge distillation by training a lightweight DistilBERT model on the CLINC150 dataset, replacing LLaMA 2 to reduce inference costs by 60% while maintaining high accuracy in Vertex AI.
• Fine-tuned the Llama-2-7B model using Reinforcement Learning from Human Feedback (RLHF), increasing alignment of invitations with desired tone and input features by 25%, enhancing customer engagement.
• Deployed Low-Rank Adaptation (LoRA) during fine-tuning, reducing trainable parameters by 88%, cutting GPU memory usage by 75%, and accelerating training time by 60%.
• Implemented model evaluation, tracking, and retraining workflows using MLflow and Airflow, reducing deployment lead time and improving model stability.

Machine Learning Research Intern (Tenure: JUN 2021– FEB 2022)

Samsung, Prism, Chennai About Prism Program at Samsung: The Samsung PRISM program (full name: Samsung PRISM (Preparing and Inspiring Student Minds)) is an industry‐academia initiative launched by Samsung’s Bangalore R&D centre (SRI-B) to engage engineering college students and faculty in real R&D projects across topics like AI, machine learning (ML), IoT and 5G. Out of 4000 students competed from my university, I was one among the 60 students they have selected for this program. My experience:

• Researched on "AI Based Reflection Scene Category Classification" in Pytorch environment and achieved an accuracy of 93% in identifying different types of reflections in real time for a dataset of 100,000 plus images.
• Applied pruning technique to reduce inference time by 60%, achieving a 93% classification accuracy, comparable to state of-the-art (SOTA) models [EfficientNet, DenseNet, MobileNetV3] for real-time reflection scene classification

PROJECTS

Enhancing RAG: unstructured data Extraction and Vector DB Evaluation (Timeline: MAY 2025 - Present)

• Developed an end-to-end framework using Docling to extract content from PDFs and images, including hierarchical metadata such as headings, titles, and table names; upserted data into a self-hosted Qdrant collection, improving PDF querying accuracy by 25% and reducing retrieval and storage costs by 100%.
• Researched, designed, and developed an evaluation framework to assess relevance of retrieved chunks across various vector databases and identify the optimal upsertion method; framework leverages bagging of metrics including cosine similarity, BERTScore, context relevance, and faithfulness.
• Currently developing a novel chunk upsertion method focused on dense and partial sparse embeddings, projected to reduce overall sparse embedding storage by 95%.

Text2Block (Timeline: OCT 2024 - OCT 2025)

• “A picture can speak a thousand words” — that’s the inspiration behind creating Text2Block.
• Designed and launched Text2Block, a GenAI application that transforms plain text into AI-driven flowchart visualizations with noise-free embedded text for enhanced clarity and improved learning. The platform aims to leverage AI for education, attracting 450 unique users and processing 3,000 requests within its first two weeks.
• Optimized system prompts for LLMs, increasing structured output accuracy by 40% using few-shot learning. Integrated Agentic LLMs in LangGraph for response evaluation and content regeneration, improving relevance to user requirements.

TAVI: Surrounding Awareness System (Google Hackathon Winner - 3rd Prize)

• Architected a multimodal accessibility system designed to enhance situational awareness for visually impaired users through real-time audio and video processing with continuous voice interaction.
• Built an end-to-end FastAPI backend with wake-word detection (pvporcupine), speech-to-text transcription (Whisper), and intent classification using GPT-based models to determine user actions (Record, General, Fallback, Tavi).
• Engineered advanced video processing pipeline integrating frame extraction (OpenCV), image captioning (Hugging Face BLIP), OCR extraction (Mistral OCR), and LLM-based summarization (ChatGroq) to generate context-rich environmental summaries, converted to speech via pyttsx3.
• Developed accessible Kivy-based mobile frontend featuring a scrolling chat interface with continuous interaction loops, displaying transcripts, text responses, and media (video previews and audio playback) optimized for visually impaired users.
• Tech Stack: Python, FastAPI, Kivy, OpenAI Whisper, Hugging Face Inference API, ChatGroq, Mistral OCR, OpenCV, pyttsx3, pvporcupine, HTTP API communication.

Deep Learning Research - Nanyang Technological University, Singapore (Timeline: AUG 2022 – AUG 2023)

• Fine-tuned CNN models, including VGGNet, ResNet, and EfficientNet, by unfreezing 20% of the neural network layers for training on the constructed dataset, resulting in a 15% improvement in test accuracy by optimizing model weights specific to the dataset.
• Implemented a novel framework for academic emotion classification using VGG-19 as a feature extractor and Multi-Layer Perceptron (MLP) as a classifier, achieving 82.73% classification accuracy on the test set after performing 5-fold cross-validation for complex emotions such as boredom and frustration (Conference Paper).
• Developed a multimodal analysis pipeline for emotion recognition in memes, achieving 89% accuracy by integrating Optical Character Recognition (OCR) for text extraction and hyperparameter tuning for image analysis in TensorFlow/Keras

EDUCATION

Master of Data Science (Timeline: AUG 2023 - MAY 2025)

Major of Study in Computer Science and Mathematics
Illinois Institute of Technology, Chicago, IL

• Graduate Pathway Scholarship Awardee ($ 10,000) awarded for Academic Merit at Illinois Institute of Technology.
• Coursework – Machine Learning, Database Organization, Generative AI, Data Preparation & Analysis, Big Data Technologies.

Bachelor of Technology in Electronics and Communication Engineering (Timeline: July 2019 - APR 2023)

Vellore Institute of Technology, Chennai, TN

• Published a research paper in fields of Deep Learning, Computer Vision, and NLP in journal Applied Sciences.

Additional Info

RESEARCH

Over-volume vehicle classification using Deep CNN models (Timeline: DEC 2021- MAY 2022)

• Collected real-time image data and performed image classification to keep a check on over-volume vehicles in the absence of human surveillance to assist the commuters in rural and terrain areas.
• Noticed an increase in performance by 12% after performing transfer learning, fine tuning and hyper-parameter tuning.
• Achieved an accuracy of 96% using EfficientNet model and published our work in applied sciences journal.

Tech Stack Used: Deep learning, Image augmentation, Computer Vision, Pytorch, Tensorflow, Team Leader, Neural Network, hyperparameter tuning, Google Colab, TPU Training, K-fold Cross Validation

Nanyang Technological University (NTU) (Timeline: AUG 2022 - AUG 2023)

• Conducted a multimodal analysis on memes, achieving a 89% accuracy in emotion recognition by integrating OCR for text extraction and deep learning for image analysis within a TensorFlow environment.
• Led research on facial emotion recognition for classroom applications, utilizing advanced computer vision techniques to enhance real-time student engagement monitoring.
• Developed and deployed a first-of-its-kind classroom emotion dataset, reducing cost of data collection by 100% through efficient web scraping, landmark detection, and facial unit mapping processes.
• Achieved a significant classification accuracy of 82.73% for complex emotions like boredom and frustration in classroom settings by integrating dataset with state-of-the-art models using TensorFlow/Keras.
• Published the research at The European Conference on Education 2023 Official Conference Proceedings.

Tech Stack Used: Deep learning, Image augmentation, Data Collection, Web Scraping, Model Building from Scratch, Computer Vision, NLP, Multi-modal training, Pytorch, Tensorflow, Team Leader, Neural Networks, hyperparameter tuning, Google Colab, TPU Training, K-fold Cross Validation, AWS sagemaker.

Further Clarifying information about myself I would like to provide to be considered in this application

Beyond what's listed on my resume, I bring a mindset that blends technical depth with practical execution. I approach problems by thinking like a systems builder—always asking how a model, pipeline, or orchestration setup will scale, integrate, and drive measurable impact.

I'm naturally curious and tend to reverse-engineer processes until I understand not just how something works, but why it behaves that way. This trait has helped me bridge gaps between data science experimentation and real-world production constraints—especially when dealing with cloud environments, large-scale data, and evolving business needs.

I also take ownership end-to-end: from designing data collection logic and validation checks to deploying APIs or RAG-based systems that are production-ready. I'm comfortable working through ambiguity, questioning assumptions, and simplifying complex systems into actionable solutions.

Lastly, I genuinely enjoy translating technical outputs into insights that make sense to stakeholders. That balance of analytical rigor and communication is something I practice deliberately, and it's become one of my defining professional habits.