📬 sudiksha1026@gmail.com | 💼 LinkedIn | Currently: Veeva Systems
Hi, I’m Sudiksha — a data science and analytics professional passionate about turning complex data into actionable business insights and building intelligent, data-driven solutions that create real-world, measurable impact. My background spans healthcare analytics, machine learning, causal inference, and generative AI, with a strong interest in applying ML and AI to solve complex business and product strategy problems.
I recently completed my Master’s in Information and Data Science from UC Berkeley while working full-time in data analytics solutions consulting at Veeva Systems. My experience spans longitudinal patient analytics, experimentation, scalable ML pipelines, recommendation systems, and AI-powered applications such as retrieval-augmented generation (RAG) systems and personalized risk engines.
I’m especially excited about opportunities in machine learning, applied AI, and data science where I can combine strong technical foundations with product thinking, user behavior insights, and business strategy to build impactful, user-focused solutions.
Capstone Project | AWS · FastAPI · React/TypeScript · PostgreSQL · Claude Haiku (Bedrock)
A mobile web application that lets consumers scan barcodes or grocery receipts to instantly check for FDA food recalls and assess ingredient-level risk.
Details
- Built a deterministic ingredient risk engine with synonym normalization and fuzzy product matching (RapidFuzz + TF-IDF cosine + Cross Encoder ensemble)
- Integrated AWS Textract for OCR receipt parsing and Claude Haiku via Bedrock for LLM-based ingredient disambiguation
- Designed and deployed full AWS infrastructure: EC2, RDS (PostgreSQL), IAM policies, SES email alerts
- Led technical evaluation, slide deck, and live demo for final presentation
Machine Learning at Scale | PySpark · MLlib · Databricks · Delta Lake · Optuna · Python
Built a distributed ML pipeline to classify U.S. flight delays across 15M+ rows of historical flight and weather data and engineering RFM (Recency, Frequency, Monetary) and graph-based features.
Details
- Class-weighted Random Forest model with 36 engineered features to handle severe label imbalance across flight and weather data
- Block-window time-series cross-validation to respect temporal dependencies and prevent data leakage across quarterly splits
- Optuna hyperparameter tuning, reducing tuning time ~40% vs. grid search; Delta Lake checkpointing cut pipeline runtime by ~60%
- Designed for real passenger impact: enabling proactive delay alerts 2–4 hours before departure to inform travel decisions
Generative AI | LangChain · Qdrant · Sentence Transformers · Cohere · Mistral · Ragas · Python
Built and evaluated an end-to-end Retrieval-Augmented Generation (RAG) pipeline to improve enterprise document search and Q&A for distinct user personas (engineering and marketing), outperforming a general-purpose LLM baseline on faithfulness, context recall, and answer relevance.
Details
- Designed a semantic retrieval pipeline using
RecursiveCharacterTextSplitterfor document chunking andmulti-qa-mpnet-base-dot-v1sentence transformer embeddings stored in a Qdrant vector database - Engineered persona-specific LLM generation prompts (Cohere + Mistral) to produce tailored responses for technical and business users, reducing hallucinations through retrieval-grounded generation
- Evaluated system performance using RAGAS metrics (faithfulness, context precision/recall, answer relevance) alongside a custom LLM-as-a-judge scorer (Qwen2.5-7B) and a weighted composite scoring framework
- Authored an executive POC report with methodology, findings, and scaling recommendations for production deployment
ML Systems Engineering | DistilBERT · FastAPI · Docker · Kubernetes · Redis · Istio
Deployed a fine-tuned DistilBERT sentiment classifier as a scalable, production-grade REST API on AWS EKS.
Details
- Containerized FastAPI service with Docker; orchestrated via Kubernetes/Kustomize with Horizontal Pod Autoscaling
- Redis caching layer to reduce redundant inference calls
- Full observability stack: Istio service mesh + Prometheus metrics + Grafana dashboards
- Load tested with Locust; documented latency and throughput benchmarks
Data Engineering | Neo4j · PostgreSQL · Redis · MongoDB · Python · Cypher · Neo4j GDS
A proof-of-concept data engineering system that models the SF Bay Area BART network as a graph database to optimize food delivery logistics — selecting high-value pickup stations, clustering delivery zones, and routing orders efficiently across 50 stations and 8,000+ customer locations.
Details
- Applied Harmonic Centrality (Neo4j GDS) on a bipartite station-customer graph with geodesic distance filtering to identify optimal BART pickup stations by customer density at 1-, 2-, and 5-mile proximity thresholds
- Implemented Louvain Modularity community detection to cluster geographically connected stations and prioritize underserved delivery zones for fleet expansion
- Designed shortest-path delivery routing across BART lines, modeling travel times and transfer costs as weighted graph edges
- Integrated PostgreSQL for structured customer and station data, Redis for real-time delivery state, and MongoDB for persistent customer profiles and order history
Statistics | R · ggplot2 · OLS Regression · sandwich/lmtest · stargazer · Zillow API Data
Descriptive regression study examining the relationship between average home values and days-to-pending-sale across 672 U.S. metropolitan regions using Zillow data (June 2024).
Details
- Applied log-log transformation to address heavy right skew in home value distribution, enabling elasticity interpretation of regression coefficients
- Implemented HC1 heteroskedasticity-consistent robust standard errors via the sandwich package to ensure valid inference on real-world housing data
- Enforced strict exploration/confirmation set split (30/70) to prevent data snooping — all modeling decisions made before touching the confirmation set
- Evaluated all large sample linear model assumptions (I.I.D., finite variance, no perfect collinearity) and documented violations with proposed remediation strategies
- Produced publication-quality regression table via stargazer comparing OLS and robust coefficient tests across model specifications
*Applied Machine Learning | Python · TensorFlow/Keras · Scikit-learn · NLTK · Pandas · NumPy Benchmarked five machine learning models — from a Logistic Regression baseline to Bidirectional LSTM and CNN — on binary sentiment classification of 100,000 real-world customer reviews spanning 45 companies, achieving a best test accuracy of 96.3%.
Details
- Preprocessed and balanced a 100K-review dataset by dropping neutral ratings and sampling 12,000 positive and 12,000 negative reviews to eliminate class bias
- Implemented and compared five architectures: Logistic Regression (baseline), Random Forest, LSTM, Bidirectional LSTM, and 1D CNN — tracking train/validation/test accuracy and loss across all
- Demonstrated that deep learning models dramatically outperform traditional ML for unstructured text: BiLSTM and CNN both exceeded 96% test accuracy vs. 66% for Logistic Regression
- Conducted ablation experiments on dropout, convolutional depth, and embedding dimensionality to assess the impact of hyperparameter choices on generalization
- Found that Bidirectional LSTM achieved the best overall test performance (96.3% accuracy, 0.111 test loss), outperforming standard LSTM by capturing both past and future token context
XAI / Computer Vision | Python · TensorFlow · Keras · Grad-CAM · Integrated Gradients · FairFace · Jekyll
An interactive research web app that applies Explainable AI (XAI) techniques to a CNN-based racial classification model, surfacing how biased training data produces discriminatory predictions — and making the model's decision-making visible to a general audience.
Details
- Trained CNN classifiers on both a balanced (FairFace, 100K+ images) and a US Census-proportioned biased dataset to directly compare fair vs. skewed model behavior
- Implemented Grad-CAM and Integrated Gradients in Keras to generate gradient heatmaps that localize which facial features drive each prediction
- Demonstrated measurable bias: the biased model misclassifies underrepresented races (e.g., Indian → Latino Hispanic) while failing to learn distinctive facial features, confirmed visually through saliency maps
- Published findings as an interactive educational website targeting public understanding of algorithmic fairness, model transparency, and AI ethics
Machine Learning · Python · Pandas · Scikit-learn · Matplotlib · Seaborn · pytz
End-to-end analysis of 2,000+ Snapchat political ads (2018–2019) to identify characteristics that drive ad reach, culminating in a fairness-evaluated classification model that predicts high vs. low view count with ~91% accuracy.
Details
- Engineered time-zone-aware features from raw billing addresses using address parsing and pytz, enabling accurate "part of day" release analysis across 115+ cities globally
- Conducted missingness analysis (NMAR/MAR/MCAR) via permutation tests with TVD,
finding
LocationTypemissingness is MAR-dependent onCity(p = 0.013) - Ran two permutation hypothesis tests confirming ad spend and release time significantly differ between high- and low-reach ads (both p = 0.0, α = 0.05)
- Built a Logistic Regression pipeline with
ColumnTransformer(OneHotEncoding + StandardScaler) tuned viaGridSearchCV, improving accuracy from 53% (baseline) to 91% (final model) - Performed fairness evaluation using false positive rate parity across view-count groups (p ≈ 0.82), confirming the final model does not disproportionately misclassify either group
Languages & Libraries
Python SQL R
ML & Stats
XGBoost scikit-learn SHAP Logistic Regression LLM BERT TF-IDF Causal Inference A/B Testing
MLOps & Cloud
AWS (EC2, EKS, Bedrock, RDS, IAM) Docker Kubernetes FastAPI Redis Prometheus Grafana Istio PostgreSQL
Data & Visualization
Pandas NumPy Matplotlib Seaborn