Conversational AI & LLM Data Scientist

DHANUSH CHOWDARY PUNATI

908-***-**** *************@*****.*** LinkedIn

SUMMARY

Data Scientist specializing in Conversational AI and Natural Language Processing (NLP), with 3+ years building virtual assistants, LLM, and machine learning systems used in production across healthcare, insurance, and education. Designs conversations end to end, from intent and entity modeling and text classification to embeddings, Transformer-based models, and fine-tuned large language models (LLMs). Builds agentic AI and multi-agent architectures with rigorous evaluation frameworks (golden datasets, LLM-as-a-Judge, error analysis), and optimizes LLM inference in production with KV caching, batching, and quantization for low latency and scalability. Also delivers classical machine learning: feature engineering, time-series forecasting, data preprocessing, and model evaluation.

CAREER HIGHLIGHTS

Launched a clinical virtual assistant that now resolves 38% of patient intake conversations without staff handoff, using LangGraph multi-agent orchestration with 45 intents and 30 entity types.

Raised intent classification accuracy from 81% to 93% by fine-tuning BERT and T5 Transformer models on 25K labeled utterances and running weekly error analysis on misclassified turns.

Cut LLM p95 response latency by 52% (2.9s to 1.4s) and GPU serving cost by 30% by serving LLaMA on vLLM with KV caching, prefix caching, continuous batching, and 4-bit quantization.

TECHNICAL SKILLS

Conversational AI Virtual Assistants, Chatbots, Conversation Design, Dialogue Management, Intent Classification, Entity Extraction/NER, Slot Filling, Fallback & Escalation Design, Error Analysis, Rasa, Dialogflow CX

NLP & LLMs Natural Language Processing (NLP), Text Classification, Embeddings, Semantic Search, Transformer Architecture (Self-Attention), BERT, T5, GPT, LLaMA, Hugging Face Transformers, Prompt Engineering, RAG

Agentic AI & Evaluation Agentic AI, Multi-Agent Architectures, LangGraph, LangChain, Tool/Function Calling, Evaluation Frameworks, Golden Datasets, LLM-as-a-Judge, RAGAS, Regression Testing

Fine-Tuning & Inference Fine-Tuning (PEFT/LoRA, QLoRA), Quantization, Inference Optimization, KV Caching, vLLM, Continuous Batching, Latency Reduction, Scalability, Model Serving

Machine Learning Feature Engineering, Time-Series Forecasting, Data Preprocessing, Model Evaluation, XGBoost, LightGBM, Scikit-learn, PyTorch, TensorFlow, Anomaly Detection, A/B Testing

Data & MLOps Python, SQL, PySpark, Apache Spark, Apache Airflow, MLflow, Docker, Kubernetes, FastAPI, CI/CD, Git, FAISS, Pinecone, PostgreSQL, MongoDB, Snowflake

Cloud AWS (SageMaker, S3, Lambda, EC2), Azure ML, GCP (Vertex AI)

WORK EXPERIENCE

AI Engineer Teladoc Health Oct 2025 – Current

Launched a clinical virtual assistant that resolves 38% of patient intake conversations without staff handoff, measured over 60K monthly conversations, by designing conversation flows, a 45-intent taxonomy, 30 entity types, and escalation paths.

Raised intent classification accuracy from 81% to 93%, measured on a held-out golden dataset, by fine-tuning BERT and T5 Transformer models on 25K labeled utterances and running weekly error analysis on misclassified turns.

Improved entity extraction F1 from 0.78 to 0.89 on medications, symptoms, and dates, measured on 3K annotated visit notes, by training NER models with PyTorch and Hugging Face and adding OCR for scanned intake forms.

Reduced multi-turn task failures by 27%, measured through conversation-level success rates, by building an agentic multi-agent architecture in LangGraph with planner, retrieval, and tool-calling agents.

Increased first-contact resolution by 19%, validated through A/B tests on 20K conversations, by designing clarifying-question and slot-filling strategies for ambiguous patient requests.

Reduced conversation drop-off by 23% by mining transcripts with embedding-based clustering to discover 12 unhandled intents and adding them to the assistant.

Built an evaluation framework that catches regressions before release, gating every prompt and model change, by creating 1.2K-case golden datasets, LLM-as-a-Judge scoring for groundedness and helpfulness, and automated CI/CD regression tests.

Cut LLM p95 latency by 52% (2.9s to 1.4s) and GPU cost by 30%, measured in production monitoring, by serving LLaMA on vLLM with KV caching, prefix caching, continuous batching, and 4-bit quantization.

Lowered hallucinated answers by 41%, measured by LLM-as-a-Judge faithfulness scores, by building RAG over clinical knowledge bases with embeddings, hybrid search, and re-ranking in FAISS and Pinecone.

Specialized LLMs for clinical intake, improving task accuracy by 15 points over the base model, by fine-tuning with PEFT/LoRA on AWS SageMaker and continuously retraining on production conversation logs.

Scaled the assistant to 3x peak traffic with 99.9% uptime by deploying FastAPI inference services with Docker, Kubernetes autoscaling, and MLflow-based model monitoring.

Forecast weekly conversation volume within 8% MAPE using time-series forecasting models (Prophet, LightGBM), enabling staffing and GPU capacity planning.

Cut annotation cost by 45% while keeping label quality above 95% agreement by using LLM-assisted pre-labeling with human review for intent and entity training data.

Reduced unsafe or off-policy responses to under 0.5% of conversations, measured by weekly audits, by adding guardrails, PII redaction, and prompt-injection checks to the agent pipeline.

Partnered with clinicians, product managers, and engineers to prioritize production improvements from error analysis, and mentored 2 junior engineers on NLP and LLM evaluation.

AI Engineer Intern Montclair State University Feb 2025 – Sep 2025

Increased student chatbot self-service resolution from 54% to 71%, measured over 8K monthly sessions, by redesigning conversation flows, intents, and fallback prompts based on transcript error analysis.

Raised intent classification F1 from 0.79 to 0.90 by fine-tuning BERT and T5 with Hugging Face Transformers and benchmarking them against RNN, LSTM, and GRU baselines in Keras.

Reduced wrong answers by 35%, measured on a 500-question golden dataset, by grounding responses in a RAG pipeline with LangChain, FAISS, and sentence embeddings.

Shortened the retraining cycle from monthly to weekly by building a feedback loop that captured misclassified intents in MongoDB, then preprocessed and relabeled them for retraining.

Cut semantic search latency by 60% (400ms to 160ms) by tuning FAISS vector index parameters and caching frequent query embeddings for real-time responses.

Reduced negative user feedback by 20% by adding sentiment analysis on chat turns that triggered human escalation for frustrated users.

Built LLM evaluation scripts in Python and Pandas that track accuracy, latency, and token cost per release, and presented results at bi-weekly reviews to prioritize fixes.

Increased intent coverage by 30% by clustering unlabeled student queries with sentence embeddings and K-Means to propose new intents for conversation design.

AI Engineer Marsh McLennan May 2023 – Jul 2024

Cut analyst document review time by 40%, measured across 15 risk analysts, by building an LLM-powered assistant with RAG (Pinecone, embeddings) and structured outputs over policy documents and risk reports.

Routed 100K+ incoming documents into 20 categories at 94% accuracy, removing manual triage, by fine-tuning BERT text classifiers on labeled policy and claims data.

Produced schema-valid JSON for 98% of extractions by applying prompt engineering, tool/function calling, and structured outputs with the OpenAI API.

Improved premium forecast accuracy by 18% (MAPE 14% to 11.5%) by building time-series forecasting models with engineered lag, seasonality, and macro features in Python and R.

Flagged 22% more anomalous claims at the same false-positive rate by developing anomaly detection models with statistical baselines and isolation forests.

Raised claims risk-scoring AUC from 0.74 to 0.82 by engineering 60+ features and tuning XGBoost and LightGBM models with cross-validated hyperparameter search.

Extracted key risk indicators from 50K+ scanned documents with 90%+ field accuracy by combining OCR, NER, and text classification pipelines.

Reduced data preparation time from days to hours by engineering ETL pipelines in Apache Spark, PySpark, and SQL that moved SQL Server policy data into Snowflake with data quality checks.

Kept models reliable in production, detecting drift within 24 hours, by maintaining a feature store, MLflow Model Registry, and model monitoring alerts.

Raised retrieval hit rate@5 from 0.68 to 0.84, measured on a 300-query golden set, by tuning document chunking, metadata filters, and embedding models for semantic search.

Served models through FastAPI and Flask REST APIs with Docker on AWS EC2, holding API latency under 300ms at peak analyst load.

Worked with product owners and risk analysts in Agile sprints to scope NLP use cases, define success metrics, and prioritize features by business value.

Drove adoption across 3 risk teams by communicating model results and limitations through Tableau and Power BI dashboards for business stakeholders.

Machine Learning Intern Codegnan IT Solutions Pvt Ltd Jan 2023 – Apr 2023

Improved baseline classification accuracy by 9% by building end-to-end ML pipelines with data preprocessing, feature engineering, hyperparameter tuning, and XGBoost and LightGBM ensembles.

Evaluated regression, classification, and clustering models with Scikit-learn using cross-validation, F1-score, and ROC-AUC, and documented model trade-offs for the team.

Automated data ingestion from multiple sources into MySQL with Python, Pandas, and SQL pipelines, removing manual weekly data pulls.

Reached 91% validation accuracy on an image classification task by training CNN deep learning models with TensorFlow and Keras.

Halved experiment setup time by exposing models through Flask REST APIs and automating training runs with Bash scripts.

Presented experiment results in Jupyter with Matplotlib and Seaborn, turning findings into 5 pipeline improvements adopted by the team.

Tracked experiments and code with Git and wrote unit tests for data and model validation, keeping results reproducible.

EDUCATION

Master's in Computer and Information Science Montclair State University Sep 2024 – May 2026

GPA: 3.70 / 4.0

Bachelor of Technology in Computer Science & Engineering (Artificial Intelligence) KKR & KSR Institute of Technology and Sciences June 2020 – Apr 2024

Graduated: May 2024 GPA: 3.35 / 4.0

Contact this candidate