AI/ML Engineer for Enterprise Systems

Sushma A

AI/ML ENGINEER

201-***-**** *************@*****.*** LinkedIn GitHub Dallas, TX, USA SUMMARY

● AI Engineer with 4+ years of experience building and deploying ML systems across finance and enterprise AI.

● Orchestrated AI pipelines integrating conversational AI and customer retention modelling, reducing financial risk and improving investment decision workflows by 25% for Fortune 500 clients.

● Architected Generative AI systems combining LangGraph multi-agent workflows with RAG (Vector DB) and LLM fine-tuning, raising retrieval relevance by 35% and cutting memo preparation time by 50% for investment research.

● Streamlined data pipelines with Apache Spark, Kafka, Airflow, and AWS, Processing over 80M+ transactions daily, automating ingestion, feature engineering, and behavioral drift analysis for model retraining.

● Operationalized AI models in cloud-native environments using Docker, Kubernetes, AWS Sagemaker AI, and FastAPI, achieving sub-200ms latency.

● Systematized workflows with MLflow, GitHub Actions, and CI/CD pipelines, supporting continuous model retraining drift detection, and versioned deployments with 99.5% deployment success rate.

● Proficient in AI/ML stacks, including PyTorch, TensorFlow, Hugging Face, Scikit-learn, BERT, LLaMA, GPT, prompt engineering, and agentic AI systems, in regulated enterprise contexts. SKILLS

Programming Languages: Python, SQL, Java, C, C++, JS, Angular, React.js, HTML/CSS Data Analytics: Pandas, NumPy, SciPy, Matplotlib, Seaborn, Plotly ML & Deep Learning: PyTorch, TensorFlow, Keras, Scikit-learn, XGBoost, Hugging Face Transformers Generative AI & NLP: GPT-4/5, BERT, LLM, RAG, GRAG, Prompt Engineering, Text Generation, Agentic AI ML Ops & Deployment: ML flow, Docker, Kubernetes, GitHub Actions, CI/CD pipelines, Canary Deployments Large Language Models: Claude 3, GPT-4, LLaMA 2/3/4, Mistral, Tokenizers, CoT Reasoning, Few-shot Learning LLM Ops & Tooling: LangChain, LangGraph, Crew AI, Pinecone, FAISS, Weaviate, Gradio, Vector Databases Cloud Platforms:

Protocols:

AWS (SageMaker, Lambda, EC2), Azure (ML Studio, Databricks), GCP (Cloud AI, Vertex) Rest API, MCP (Model Context Protocol), Fast API, Open AI APIs Databases: MySQL, PostgreSQL, SQL Server, Vector DB, Chroma DB, Elasticsearch Big Data & Pipelines: Apache Spark, Kafka, Airflow, Feature Engineering, Data Cleaning, Normalization Model Evaluation: SHAP, ROC-AUC, F1 Score, Precision, Recall, Cross-Validation, Statistical Testing, BLEU, ROUGE

Software Development:

Operating System:

OOP principles, Multi-threading, Design patterns, Agile/Scrum methodologies, Eclipse, Visual Studio Code, Version Control (Git, GitHub, GitLab, GitHub Actions) Windows, Linux (Ubuntu, Red Hat)

Model Development:

Supervised/Unsupervised learning, Time series forecasting, Anomaly Detection, Reinforcement learning, Semantic segmentation (OpenCV)

EXPERIENCE

AI/ML Engineer FIDELITY INVESTMENTS, TX, USA Aug 2024 – Present

● Built GenAI pipeline automating Chief Investment Office workflows, integrating portfolio optimization, real-time market intelligence, and compliance monitoring, serving 500+ investment professionals.

● Constructed predictive portfolio allocation models using XGBoost and Scikit-learn, incorporating historical returns, and improving risk-adjusted asset allocation strategies by 12% across live trading.

● Orchestrated data pipelines using Apache Kafka, AWS Glue, and Athena, Processing market feeds, financial news, and regulatory updates with 99.8% into structured event streams for downstream GenAI agents.

● Formulated feature engineering framework for financial signals, generating time-series windows, cross-sector risk indicators, supporting both traditional ML models and LLM decision agents.

● Executed multi-agent orchestration with LangGraph and MCP (Model Context Protocol) using LLaMA 3 8B for research, compliance, and reporting agents; applied semantic routing and task-specific prompts to cut manual memo creation time by 50%.

● Optimized large language models (FinGPT, Palmyra) using QLoRA and PEFT (Parameter Efficient Fine-tuning), Optimizing financial summarization and document classification while reducing GPU memory by 60%.

● Crafted Prompt Engineering system with few-shot learning, Chain-of-thought prompting, and context construction for personalized market analysis generation and client-ready investment narratives.

● Established a RAG (Retrieval-Augmented Generation) pipeline leveraging FAISS and Redis for indexing of earnings reports, analyst notes, and regulatory filings, with production indexing on Azure Vector DB (Azure AI Search) using HNSW + hybrid BM25/semantic reranking, achieving 35% improvements in retrieval relevance.

● Launched LLM systems and ML models using AWS SageMaker, and Docker, establishing real-time endpoints with CI/CD, reducing deployment cycle time by 30% and ensuring version control.

● Instituted MLOps and LLMOps infrastructure with MLflow, supporting automated model retraining, continual evaluation, drift detection, and reproducible experiment tracking in production environments.

● Pioneered the development of LLM evaluations, implementing metrics like BLEU, ROUGE, perplexity, and financial relevance scoring, wired as GitHub Actions quality gates on PRs/releases to block regressions and ensure consistent output quality. AI/ML Engineer Capgemini May 2020 – Jun 2023

● Delivered a fraud detection system using PyTorch and deep neural networks, reducing false positives by 31% while improving anomaly detection sensitivity in high-volume transactional environments.

● Calibrated classification models resolving severe class imbalance via SMOTE and financial KPI incorporation, boosting recall across 5+ financial institutions' data.

● Synthesized NLP pipelines with Hugging Face Transformers (BERT) and custom tokenization to extract risk entities, intent signals, and customer sentiment from KYC documents, improving monitoring time by 30%.

● Containerized ML inference services with Docker, and AWS SageMaker Endpoints, achieving a 38% reduction in model deployment time and enabling A/B testing and canary rollouts in production.

● Integrated fraud and customer behavior scoring using FastAPI, integrating real-time predictions for both customer NPS segmentation (promoter, passive, detractor) and fraud detection with sub-200 ms latency on Lambda, ingesting events from Kafka and using Spark windowed aggregations; end-to-end orchestration via Airflow.

● Accelerated hyperparameter optimization using Optuna, enhancing model convergence speed and improving F1-score compared to random by 10%, avoiding overfitting legacy patterns.

● Visualized explainable AI dashboards with Plotly and Seaborn, integrating SHAP values, feature attributions, and fraud signal analytics, reducing compliance investigation time by 19% and enhancing audit trail transparency.

● Standardized the ML lifecycle with MLflow and GitHub Actions, for model retraining, data versioning, and drift detection, cutting manual intervention by 25% and ensuring deployments across staging and production.

● Validated model pipelines using ROC-AUC, cross-validation, and statistical tests with Statsmodels, generating regulatory- grade risk scores with high interpretability, meeting strict compliance and audit standards. EDUCATION

MASTER OF SCIENCE in Computer Science, University North Texas, TX, US PROJECTS

FSDD - Spoken Digit Classifier (Audio) GitHub Demo Video

● Achieved 94.1% test accuracy on the Free-Spoken Digit Dataset (FSDD) with a stratified 70/15/15 split and fixed seed; added confusion-matrix diagnostics and per-class precision/recall.

● Engineered a portable data loader uses Hugging Face Datasets when available, auto-fallback to the official FSDD GitHub ZIP on Windows, ensuring the project runs on any laptop.

● Implemented robust audio preprocessing: resample to 8 kHz, mono, pad/trim to 1s, normalize, and extract MFCCs-keeps features tiny and stable across recording conditions.

● Shipped a Streamlit app with Upload and Record tabs, mic input via browser, end-to-end latency 8–15 ms on CPU, enabling real-time predictions.

● Productionized the pipeline as a single joblib (StandardScaler LogisticRegression)-model artifact is 6 KB, easy to version and deploy.

● Operationalized evaluation & monitoring: after each training run, a CI job (GitHub Actions) generates and stores the confusion matrix and classification report as build artifacts, applies quality gates (per-class precision/recall), and blocks deploy on regression; results are surfaced for traceability in the repo/Streamlit app.

● Practiced LLM-assisted development: used targeted prompts for architecture decisions, HF-fallback logic, evaluation narration, and UI polish.

Contact this candidate